research article a hybrid multiobjective differential
TRANSCRIPT
Hindawi Publishing CorporationJournal of Applied MathematicsVolume 2013 Article ID 841780 15 pageshttpdxdoiorg1011552013841780
Research ArticleA Hybrid Multiobjective Differential EvolutionAlgorithm and Its Application to the Optimization ofGrinding and Classification
Yalin Wang1 Xiaofang Chen1 Weihua Gui1 Chunhua Yang1
Lou Caccetta2 and Honglei Xu2
1 School of Information Science and Engineering Central South University Changsha 410083 China2Department of Mathematics amp Statistics Curtin University Perth WA 6845 Australia
Correspondence should be addressed to Xiaofang Chen xiaofangchencsueducn
Received 19 July 2013 Accepted 5 September 2013
Academic Editor Dewei Li
Copyright copy 2013 Yalin Wang et al This is an open access article distributed under the Creative Commons Attribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited
The grinding-classification is the prerequisite process for full recovery of the nonrenewable minerals with both production qualityand quantity objectives concerned Its natural formulation is a constrainedmultiobjective optimization problem of complex expres-sion since the process is composed of one grinding machine and two classification machines In this paper a hybrid differentialevolution (DE) algorithm with multi-population is proposed Some infeasible solutions with better performance are allowed tobe saved and they participate randomly in the evolution In order to exploit the meaningful infeasible solutions a functionallypartitionedmulti-populationmechanism is designed to find an optimal solution from all possible directions Meanwhile a simplexmethod for local search is inserted into the evolution process to enhance the searching strategy in the optimization processSimulation results from the test of some benchmark problems indicate that the proposed algorithm tends to converge quicklyand effectively to the Pareto frontier with better distribution Finally the proposed algorithm is applied to solve a multiobjectiveoptimization model of a grinding and classification process Based on the technique for order performance by similarity to idealsolution (TOPSIS) the satisfactory solution is obtained by using a decision-making method for multiple attributes
1 Introduction
Grinding-classification is an important prerequisite processfor most mineral processing plants The grinding processreduces the particle size of raw ores and is usually followed byclassification to separate them into different sizes Grinding-classification operation is required to produce pulp withsuitable concentration and fineness for flotation The pulpquality will directly influence the subsequent flotation effi-ciency and recovery of valuable metals from tailings In orderto improve economic efficiency and energy consumption theprocess optimization objectives include product quality andoutput yields Under certain mineral source conditions theobjectives are decided by a series of operation variables suchas the solid flow of feed ore to ball mill the steel ball fillingrate and the flow rates of water added to the first and the
second classifier recycles To solve the optimization modelof productsrsquo output and quality in the grinding-classificationprocess is of great significance to improve the technicaland economic specifications and it has been a continuousendeavor of the scientists and engineers [1ndash3]
Grinding-classification is an energy-intensive processinfluenced bymany interacting factorswithmutual restraintsThe goals of grinding-classification optimization problemare decided by multiple constrained input control variablesof nonlinear relationships So the optimization model ofgrinding-classification operation is a complex constrainedmultiobjective optimization problem (CMOP) Generallyconstrained multiobjective problems are so difficult to besolved that the constraint handling techniques and multi-objective optimization methods need to be combined foroptimization
2 Journal of Applied Mathematics
Multiobjective optimization problems (MOPs) in thecase of traditional optimization methods are often handledby aggregating multiple objectives into a single scalar objec-tive through weighting factors MOPs have a set of equallygood (nondominating) solutions instead of a single onecalled a Pareto optimum which was introduced by Edge-worth in 1881 [4] and later generalized by Pareto in 1896[5] The practical MOPs are often implicated in series ofequations functions or procedures with complicated con-straints Therefore the evolutionary algorithms are attractiveapproaches for low requirements onmathematical expression[6] Since the mid 1980s there has been a growing interestin solving MOPs using evolutionary approaches [7ndash10]One of the most successful evolutionary algorithms for theoptimization of continuous space functions is the differentialevolution (DE) [11] DE is simple and efficiently converges tothe global optimum in most cases [12 13] Its efficiency hasbeen proven [14] in many application fields such as patternrecognition [15] and mechanical engineering [16]
There have been many improvements for DE to solveMOPs Abbass [17] firstly provided a Pareto DE (PDE) algo-rithm for MOPs in which DE was employed to create newsolutions and only the nondominated solutions were kept asthe basis for the next generation Madavan [18] developed aPareto differential evolution approach (PDEA) in which newsolutions were created by DE and kept in an auxiliary popu-lation Xue et al [19] introduced multiobjective differentialevolution (MODE) and used Pareto-based ranking assign-ment and crowding distancemetric but in a differentmannerfrom PDEA Robic and Filipi [20] also adopting Pareto-based ranking assignment and crowding distance metricdeveloped a DE for multiobjective optimization (DEMO)with a different population update strategy and achievedgood results Huang et al [21] extended the self-adaptive DE(SADE) to solve MOPs by a so called multiobjective self-adaptive DE (MOSADE) They further extended MOSADEby using objectivewise learning strategies [22] Adeyemo andOtieno [23] provided multiobjective differential evolutionalgorithm (MDEA) In MDEA a new solution was generatedby DE variant and compared with target solution If itdominates the target solution then it was added to the newpopulation otherwise a target solution was added
On the other hand single-objective constrained opti-mization problems have been studied intensively in the pastyears [24ndash28] Different constraint handling techniques havebeen proposed to solve constrained optimization problemsMichalewicz and Schoenauer [29] divided constraints han-dling methods used in evolutionary algorithms into fourcategories preserving feasibility of solutions penalty func-tions separating the feasible and infeasible solutions andhybrid methods The differences among these methods arehow to deal with the infeasible individuals throughout thesearch phases Currently the penalty functionmethod ismostwidely used and this algorithm strongly depends on thechoice of the penalty parameter
Although the multiobjective optimization and the con-straint handling problem have received lots of contributionrespectively the CMOPs are still difficult to be solved inpractice Coello and Christiansen [30] proposed a simple
approach to solve CMOPs by ignoring any solution that vio-lates any of the assigned constraints Deb et al [8] proposeda constrained multiobjective algorithm based on the conceptof constrained domination which is also known as superi-ority of the feasible solution Woldesenbet et al [31] intro-duced a constraint handling technique based on adaptivepenalty functions and distance measures by extending thecorresponding version for the single-objective constrainedoptimization
In the MOP of grinding and classification process thedefinitions of Pareto solutions Pareto frontier and Paretodominance are in consistency with the classic definitionsClearly the Pareto frontier is amapping of the Pareto-optimalsolutions to the objective space In the minimization sensegeneral constrained MOPs can be formulated as follows
min 119865 (119883) = min [1198911(119883) 119891
2(119883) 119891
119903(119883)]
st 119892119894(119883) le 0 (119894 = 1 2 119901)
ℎ119895(119883) = 0 (119895 = 119901 + 1 119902)
119909119896isin [119909119896min 119909119896max] (119896 = 1 2 119899)
(1)
where 119865(119883) is the objective vector 119883 = (1199091 119909
119899) isin R119899 is
a parameter vector 119892119894(119883) is the 119894th inequality constraint
and ℎ119895(119883) is the 119895th equality constraint 119909
119896min and 119909119896max
are respectively the lower and upper bounds of the decisionvariable 119909
119896
In this paper based on the specific industrial backgroundof continuous bauxite grinding-classification operation anew hybrid DE algorithm is proposed to solve complex con-strained multiobjective optimization problems Firstly ahybrid DE algorithm for MOPs with simplex method (SM-DEMO) is designed to overcome the problems of globalperformance degradation and being trapped in local opti-mum Then for the MOPs with complicated constraints theproposed algorithm is formed by combining SM-DEMO andfunctional partitioned multi-population In this method theconstruction of penalty functions is not required and themeaningful infeasible solutions are fully utilized
The remainder of the paper is structured as followsSection 2 describes the SM-DEMO algorithm for uncon-strained cases The proposed algorithm of multipopulationfor constrained MOPs is given in Section 3 with verificationof performance by benchmark testing results Section 4describes the model of productsrsquo output and quality in thegrinding-classification process in detail and the applicationof the proposed algorithm in the optimizationmodel Finallythe conclusions based on the present study are drawn inSection 5
2 SM-DEMO Algorithm forUnconstrained MOPs
In order to efficiently solve multiobjective optimizationproblem and find the approximately complete and nearop-timal Pareto frontier we proposed a hybrid DE algorithmfor unconstrained multiobjective optimization with simplexmethod
The differential evolution with initialization crossoverand selection as in usual genetic algorithms uses a pertur-bation of two members as the mutation operator to produce
Journal of Applied Mathematics 3
a new individualThemutation operator of the DE algorithmis described as follows
Considering each target individual 119909119866119894 in the 119866th gener-
ation of size119873119901 a mutant individual 119909119866+1119894
is defined by
119909119866+1
119894= 119909119866
1199033+ 119865 (119909
119866
1199031minus 119909119866
1199032) (2)
where indexes 1199031 1199032 and 119903
3represent mutually different
integers that are different from 119894 and that are randomlygenerated over [1119873119901] and 119865 is the scaling factor
The simplex method proposed by Spendley Hext andHimsworth and later refined by Nelder andMead (NM) [32]is a derivative-free line-search method that is particularlydesigned for traditional unconstrained minimization scenar-ios Clearly NM method can be deemed as a direct line-search method of the steepest descent kind The ingredientsof the replacement process consist of four basic operationsreflection expansion contraction and shrinkage Throughthese operations the simplex can improve itself and approxi-mate to a local optimum point sequentially Furthermore thesimplex can vary its shape size and orientation to adapt itselfto the local contour of the objective function
21 Main Strategy of SM-DEMO The SM-DEMO algorithmis improved by the following three points compared with DE
211 Modified Selection Operation After traditional DE evo-lution the individual 119906119866+1
119894119895may violate the boundary con-
straints 119909max119894119895
and 119909min119894119895
119906119866+1119894119895
is replaced by new individual119908119866+1
119894119895being adjusted as follows
119908119866+1
119894119895
=
119909max119894119895
+ rand () lowast (119909max119894119895
minus 119906119866+1
119894119895) if (119906119866+1
119894119895gt 119909
max119894119895
)
119909min119894119895
+ rand () lowast (119909min119894119895
minus 119906119866+1
119894119895) if (119906119866+1
119894119895lt 119909
min119894119895
)
119906119866+1
119894119895 otherwise
(3)
The new population is combined with the existing parentpopulation to form a new set 119872119892 of bigger size than 119873119901A nondominated ranking of 119872119892 is performed and the 119873119901best individuals are selected This approach allows a globalnondomination checking between both the parent and thenew generation rather than only in the new generation asis done in other approaches whereas it requires additionalcomputational cost in sorting the combined
212 Nondominated Ranking Based on Euclidean DistanceThe solutions within each nondominated frontier that residein the less crowded region in the frontier are assigned a higherrank as the NSGA-II algorithm [8] developed by Deb et alindicated The crowding distance of the 119894th solution in itsfrontier (marked with solid circles) is the average side lengthof the cuboids (shown with a dashed box in Figure 1(a))The crowding-distance computation requires sorting thepopulation according to each objective function value inascending order of magnitude As shown in Figure 1 119860 and119862 are two solutions near 119861 in the same rank and 120590
0(119861) is the
crowding distance of the119861th solution traditionally calculatedas follows
1205900(119861) =
119899
sum
119895=1
10038161003816100381610038161003816119891119895(119860) minus 119891
119895(119862)
10038161003816100381610038161003816 (4)
where 119891119895(119860) 119891
119895(119862) are the objective vectors For each
objective function the boundary solutions (solutions withthe smallest and the largest function values) are assigned aninfinite distance value
A crowding-distance metric is used to estimate thedensity of solutions surrounding a particular solution in thepopulation and is obtained from the average distance of thetwo solutions on either side of the solution along each of theobjectives As shown in Figure 1119860119861119862 are the individuals ofthe generation on the same frontier and we easily know thatthe density in Figure 1(a) is better than that in Figure 1(b) Ifwe use (4) to calculate the crowding distance of 119862 we onlyknow that in Figure 1(a) it is better than in Figure 1(b) thecrowding distance of119862 in Figures 1(a) and 1(c) is equal whichis against the knowledge
To distinguish the mentioned situations we proposean improved crowding-distance metric based on Euclideandistance 119872 is the center point of the line 119860119861 119891
119895is the 119895th
objective vector and the crowding distance 120590(119861) is defined asfollows
120590 (119861) = |119860119862| minus |119861119872|
= radic
119899
sum
119895=1
[119891119895(119860) minus 119891
119895(119862)]2
minus radic
119899
sum
119895=1
119891119895(119861) minus
[119891119895(119860) + 119891
119895(119862)]
2
2
(5)
The crowded-comparison operator guides the selectionprocess at the various stages of the algorithm toward auniformly distributed Pareto-optimal frontier To carry outthe comparison we assume that every individual in thepopulation has two attributes (1) nondomination rank 119894rankand (2) crowding distance 119894distance Then a partial order isdefined as 119894 ≺ 119895 If 119894 ≺ 119895 that is between two solutionswith different nondomination ranks we prefer the solutionwith the lower (better) rank namely 119894rank ≺ 119895rank Otherwiseif both solutions belong to the same frontier that is 119894rank =
119895rank then we prefer the solution that is located in a lessercrowded region that is 119894distance ≻ 119895distance
213 Simplex Method for Local Search The simplex methodfor local search is mixed in the evolution process to enhancethe searching strategy in the optimization process The goalof integrating NM simplex method [32] and DE is to enrichthe population diversity and avoid being trapped in localminimum We apply NM simplex method operator to thepresent population when the number of iterations is greaterthan a particular value (like 119866max2) The individuals thatachieved the single extreme value in each objective functionaremarked as the initial vertex points of simplexmethodThe
4 Journal of Applied Mathematics
A
B
C
f1
f2
(a)
A
B
C
f1
f2
(b)
A
B
C
f1
f2
(c)
A
B
C
M
f1
f2
(d)
Figure 1 Crowding-distance diagram
present population is updated according to simplex methoduntil the terminal conditions are satisfied
The computation steps of the algorithm are included inSection 32
22 Evaluation Criteria Unlike the single-objective opti-mization it is more complicated for solution quality eval-uation in the case of multiobjective optimization Many ofthe suggested methods can be summarized in two typesOne is to evaluate the convergence degree by computing theproximity between the solution frontier and the actual Paretofrontier The other is to evaluate the distribution degree ofthe solutions in objective space by computing the distancesamong the individuals Here we choose both methods toevaluate the performance of the SM-DEMO algorithm
(1) Convergence Evaluation Deb et al [8] proposed thismethod in 2002 It is described as follows
120574 =1
119876(
119876
sum
119894=1
min 1003817100381710038171003817119875lowastminus 119875FT
1003817100381710038171003817) (6)
where 120574 is the extent of convergence to a known of Pareto-optimal set 119875lowast is the obtained nondomination Pareto fron-tier 119875FT is the real nondomination Pareto frontier 119875lowast minus119875FTis the Euclidean distance of 119875lowast with 119875FT and119876 is the numberof obtained solutions
(2) Distribution Degree EvaluationThenonuniformity in thedistribution is measured by SP as follows
SP = radic 1
(119876 minus 1)
119876
sum
119894=1
(119889 minus 119889119894)2
(7)
where 119889119894is the Euclidean distance among consecutive solu-
tions in the obtained nondominated set of solutions andparameter 119889 is the average distance
23 Experimental Studies Four well-known benchmark testfunctions [33] are used here to compare the performance ofSM-DEMO with NSGA-II DEMOParent These four prob-lems are called ZDT2 ZDT3 ZDT4 and ZDT6 each has twoobjective functions We describe them in Table 1
Journal of Applied Mathematics 5
Table 1 Test problems
Test problems Objective functionsmin119865(119883) = min[119891
1(119883) 119891
2(119883)]
Range of variable
ZDT21198911(119883) = 119909
1 1198912(119883) = 119892(119883)(1 minus (
1198911
119892(119883))
2
)
119892 (119883) = 1 + 9
119899
sum
119894=2
119909119894
119899 minus 1
119899 = 30
0 le 119909119894le 1
ZDT31198911(119883) = 119909
1 1198912(119883) = 119892(1 minus radic(
1198911
119892) minus (
1198911
119892) sin (10120587119891
1))
119892 (119883) = 1 + 9
119899
sum
119894=2
119909119894
119899 minus 1
119899 = 30
0 le 119909119894le 1
ZDT41198911(119883) = 119909
1 1198912(119883) = 119892(1 minus radic(
1198911
119892))
119892 (119883) = 1 + 10 (119899 minus 1) +
119899
sum
119894=2
(1199092
119894minus 10 cos (4120587119909
119894))
119899 = 10
0 le 1199091le 1
minus5 le 119909119894le 5 119894 = 2 119899
ZDT61198911(119883) = 1 minus exp (minus4119909
1) sin6 (6120587119909
1) 1198912(119883) = 119892(1 minus (
1198911
119892)
2
)
119892 (119883) = 1 + 9
119899
sum
119894=2
119909119894
(119899 minus 1)025
119899 = 10
0 le 119909119894le 1
Table 2 The performance results of the each algorithm on the test function
Test function Algorithm 120574 SP
ZDT2DEMOParent 0005120 plusmn 0000312 0000630 plusmn 0000010
NSGA-2 0007120 plusmn 0000413 0000540 plusmn 0000940
SM-DEMO 0004013 plusmn 0000230 0000423 plusmn 0000011
ZDT3DEMOParent 0009704 plusmn 0000027 0007512 plusmn 0000165
NSGA-2 0014067 plusmn 0000059 0006540 plusmn 0000124
SM-DEMO 0004704 plusmn 0000003 0004450 plusmn 0000153
ZDT4DEMOParent 2009704 plusmn 0901164 0011031 plusmn 0001104
NSGA-2 3144067 plusmn 2100740 0010122 plusmn 0000072
SM-DEMO 0874001 plusmn 0014323 0008721 plusmn 0000159
ZDT6DEMOParent 0649704 plusmn 0004912 0104520 plusmn 0015486
NSGA-2 1014067 plusmn 0010421 0007942 plusmn 0000105
SM-DEMO 0007750 plusmn 0000083 0002014 plusmn 0000117
The simulation is carried out under the environment ofIntel Pentium 4 CPU 306GHz 512MB memory WindowsXP Professional Matlab71 Initialization parameters are setas follows population size119873119901 = 100 scaling factor 119865 = 08cross rate 119862
119877= 06 maximum evolution generation 119866max =
250 and number of SM evolution iterations 119866SM = 100All of the three algorithms are real coded with equal
population size and equal maximum evolution generationEach algorithm independently runs 20 times for each testfunction Because we cannot get the real Pareto-optimal setwewill take 60 Pareto-optimal solutions obtained by the threealgorithms as a true Pareto-optimal solution set
We evaluated the algorithms based on the two perfor-mance indexes 120574 and SP Table 2 shows themean and varianceof 120574 and SP using three algorithms SM-DEMO NSGA-II and DEMOParent We can learn from Table 2 for theZDT2 function that all of the three algorithms have a good
performance while the SM-DEMO is slightly better than theother two algorithms In terms of convergences for ZDT3ZDT4 and ZDT6 which are more complex than ZDT2 SM-DEMO is significantly better thanDEMOParent andNSGA-II
Figure 2 shows a random running of SM-DEMO algo-rithm It is clear that SM-DEMO algorithm can produce agood approximation and a uniform distribution
3 Proposed Hybrid Algorithm for CMOP
The space of constrained multiobjective optimization prob-lem can be divided into the feasible solution space and theinfeasible solution space as shown in Figure 3 where 119878 isthe search space Ω is the feasible solution space and 119885 isthe infeasible solution space 119909
119894(119894 = 1 2 3 4) is the feasible
solution and 119910119894(119894 = 1 2 3 4) is the infeasible solution
6 Journal of Applied Mathematics
SM-DEMOPareto frontier
0 01 02 03 04 05 06 07 08 09 1f1
0
01
02
03
04
05
06
07
08
09
1f2
(a) ZDT2
0
02
04
06
08
1
minus08
minus06
minus04
minus02
f2
SM-DEMOPareto frontier
0 01 02 03 04 05 06 07 08 09 1f1
(b) ZDT3
SM-DEMOPareto frontier
0 01 02 03 04 05 06 07 08 09 1f1
0
01
02
03
04
05
06
07
08
09
1
f2
(c) ZDT4
SM-DEMO
0
01
02
03
04
05
06
07
08
09
1
Pareto frontier
02 03 04 05 06 07 08 09 1f1
f2
(d) ZDT6
Figure 2 SM-DEMO simulation curve
Assume that 119909lowast is the global optimal solution and 1199101is
the closest one to 119909lowast If the infeasible population 1199101is not
excluded by the evolution algorithm it is permitted to exploreboundary regions from new directions where the optimumis likely to be found
31 General Idea of the Proposed Algorithm Researchershave gradually realized the merit of infeasible solutions insearching for the global optimum in the feasible region Someinfeasible solutions with better performance are allowed to besaved Farmani et al [34] formulated a method to ensure thatinfeasible solutions with a slight violation become feasible inevolution Based on the constraints processing approach ofmultiobjective optimization problems the proposed hybridDE algorithm avoids constructing penalty function anddeleting meaningful infeasible solutions directly
Here the proposed algorithm will produce multiplegroups of functional partitions which include an evolution-ary population 119875119892 of size 119873119901 an intermediate population119872119892 to save feasible individuals an intermediate population119878119892 to save infeasible individuals a population 119875119891 to save theoptimal feasible solution found in the search process and apopulation 119875119888 to save the optimal infeasible solution Therelationship of multi-population is shown in Figure 4
With the description of (1) equality constraints are alwaystransformed into inequality constraints as |ℎ
119895(119883)| minus 120575 le 0
where 119895 = 119901 + 1 119902 and 120575 is a positive tolerance valueTo evaluate the infeasible solution the degree of constraintviolation of individual 119883 on the 119895th constraint is calculatedas follows
119881119894(119883) =
max 0 119892119894(119883) (119894 = 1 2 119901)
max 0 10038161003816100381610038161003816ℎ119895 (119883)10038161003816100381610038161003816minus 120575 (119895 = 119901 + 1 119902)
(8)
Journal of Applied Mathematics 7
Feasible solutionspace Ω
Search space S
x2
x1y1
y4
y3
y2 x3
x4
xlowast
Infeasible solutionspace Z
Figure 3 Distribution diagram of search space
Evolutionarypopulation
Pg
Intermediatepopulation
Mg
Optimal Randomly selected
Intermediatepopulation
SgInfeasible
Feasibleindividuals
individuals
Randomlyselected
Evolution
g = g + 1
g = g + 1
g = g + 1Randomly selected
feasiblesolution Pf
Optimal infeasible
solution Pc
Figure 4The relationship diagram ofmultipopulation demonstrat-ing
The final constraint violation of each individual in thepopulation can be obtained by calculating the mean of thenormalized constraint violations
In order to take advantage of the infeasible solutionswith better performance we proposed the following adaptivedifferential mutation operator to generate individual varia-tion learning from the mutation operators in DErand-to-best1bin according to rules defined by Price et al [11]Considering each individual vector 119909119866
119894 a mutant individual
119909119866+1
119894is defined by
119909119866+1
119894
= 119909119866
119894+ 1198651sdot (119909119866
1198911minus 119909119866
1199031)+ 1198652sdot (119909119866
1198912minus 119909119866
1199032) 119877
119862ge rand ( )
119909119866
119894+ 1198651sdot (119910119866
119894minus 119909119866
1199031)+ 1198652sdot (119909119866
1198911minus 119909119866
1199032) 119877
119862lt rand ( )
(9)
where 1199031and 1199032represent different integers and also different
from 119894 randomly generated over [1119873119901] 119865 is the scalingfactor 119909119866
119891119894(119894 = 1 2 119899) is randomly generated from 119875119891
119910119866
119894(119894 = 1 2 119899) is randomly generated from 119875119888 and 119877
119862is
the mutation factor as follows
119877119862= 1198771198880sdot120574119866+ const
120574119866minus1 + const (10)
where 1198771198880is the initial value of the variability factor const is
a small constant to ensure that the fractional is meaningfuland 120574119866 is defined as follows
120574119866=the number of infeasible solutions in 119875119892
119873119901 (11)
32 Framework of the Proposed Algorithm The proposedalgorithm is described as follows
Step 1 (initialization) Generate the population 119875119892 119875119891 and119875119888 of size 119873119901 119873119875
1 and 119873119875
2 Set the value of 119862
119877(crossover
probability)119866max (the number of function evaluations)119866SM(the iterative number of evolution by NM simplex method)119892 = 1 (the current generation number) and positive controlparameter for scaling the difference vectors 119865
1 1198652 Randomly
generate the parent population 119875119892 from the decision spaceSet the 119875119891 and 119875119888 and let the intermediate populations 119878119892and119872119892 be empty
Step 2 (DE reproduction) By (3) and (9) for mutationcrossover and selection an offspring 119878119892 is created Judge theconstraints of all individuals in 119875119892 In accordance with (8)we first calculate constraint violation degree 119881
119894(119883) of all of
the individuals If 119881119894(119883) = 0 the solution is feasible and
preserved to the intermediate set 119872119892 if 119881119894(119883) gt 0 the
solution is infeasible and preserved to the intermediate set 119878119892
Step 3 (simplex method) Apply NM simplex method oper-ator to the present population if 119892 ge 119866max2 Updatethe present population 119872119892 when the number of iterationsexceeds maximum iterations
Step 4 (119875119891 construction) Rank chromosomes in 119872119892 basedon (5) and generate the elitist population 119875119891 (the size is119873119901)from the ranked population119872119892
Step 5 (119875119888 construction) Add the chromosomes in 119878119892 withslight constraint violations to the 119875119888
Step 6 (mixing the population) Combine 119878119892with the existingparent population to form a new set119872119892 Remove the dupli-cate individuals in119872119892 and the existing parent population
Step 7 (evolution) Randomly choose chromosomes from119875119888 119875119892 and 119875119891 Use the adaptive differential mutation anduniform discrete crossover to obtain the offspring population119875119892 + 1
Step 8 (termination) If the stopping criterion ismet stop andoutput the best solution else go to Step 2
33 Experimental Study In this section we choose threeproblems CTP TNK and BNH as shown in Table 3 to testthe proposed method and compare the method with thecurrent CNSGA-II [35]
8 Journal of Applied Mathematics
Table 3 Test functions
Testfunction
Objective functionmin119865(119883) = min[119891
1(119883) 119891
2(119883)]
Constraints Range ofvariable
CTP
1198911(119883) = 119909
1
1198912(119883) = exp(minus
1198911(119883)
119888 (119883))
times41 +
5
sum
119894=2
[119909119894
2minus 10 cos (2120587119909
119894)]
1198921(119883) = cos (120579) (119891
2(119883) minus 119890) minus sin (120579) 119891
1(119883)
1198922(119883) = 119886
1003816100381610038161003816sin 119887120587 [sin (120579) (1198912 (119883) minus 119890)+ cos (120579) 119891
1(119883)]119888
10038161003816100381610038161003816
119889
1198921(119883) ge 119892
2(119883)
0 le 1199091le 1
minus5 le 119909119894le 5
119894 = 2 3 4 5
BNH1198911(119883) = 4119909
1
2+ 41199092
2
1198912(119883) = (119909
1minus 5)2+ (1199092minus 5)2
1198921(119883) = (119909
1minus 5)2+ 1199092
2minus 25
1198922(119883) = minus(119909
1minus 8)2+ (1199092+ 3) + 77
1198921(119883) le 0 119892
2(119883) le 0
0 le 1199091le 5
0 le 1199092le 3
TNK1198911(119883) = 119909
1
1198912(119883) = 119909
2
1198921(119883) = minus119909
1
2minus 1199092
2+ 1
+01 cos(16 arctan(11990911199092
))
1198922(119883) = (119909
1minus 05)
2+ (1199092minus 05)
1198921(119883) le 0 119892
2(119883) minus 05 le 0
0 le 119909119894le 120587
119894 = 1 2
20
15
10
5
00
Feasible solutionInfeasible solution
02 04 06 08 10f1
f2
Pareto-optimal frontier
(a) CTP1
20
15
10
05
0002 04 06 08 10
Feasible solutionInfeasible solution
Pareto-optimal frontier
f1
f2
(b) CTP2
Figure 5 CTP solution space
For CTP problem there are the six parameters 120579 119886 119887 119888119889 and 119890 that must be chosen in a way so that a portion ofthe unconstrained Pareto-optimal region is infeasible Eachconstraint is an implicit non-linear function of decisionvariables Thus it may be difficult to find a number ofsolutions on a non-linear constraint boundary We take twosets of values for six parameters in CTP problem which aredetermined as (1) CTP1 120579 = 01120587 119886 = 40 119887 = 05 119888 = 1119889 = 2 and 119890 = minus2 (2) CTP2 120579 = minus02120587 119886 = 02 119887 = 10119888 = 1 119889 = 6 and 119890 = 1 The Pareto frontiers the feasiblesolution spaces and the infeasible solution spaces are shownin Figure 5
The parameters are initialized as followsThe size of pop-ulation 119875119892 is119873119901 = 200 size of 119875119891 is119873119875
1= 150 size of 119875119888 is
1198731198752= 10 scaling factors 119865
1and 119865
2are randomly generated
within [05 1] cross rate is 119862119877= 06 maximum evolution
generation is 119866max = 300 and number of SM evolutioniterations is 119866SM = 100 All of the proposed algorithmsand CNSGA-II are real coded with equal population size andequal maximum evolution generation Each algorithm runs20 times independently for each test function
Figure 6 shows the result of a random running of theproposed algorithm and NSGA-II the smooth curve ldquomdashrdquorepresents the Pareto frontier and ldquoXrdquo stands for the solutionachieved by the proposed algorithm or NSGA-II
It is obvious that the proposed algorithm returns abetter approximation to the true Pareto-optimal frontie anda distribution of higher uniformity We also evaluated the
Journal of Applied Mathematics 9
40
35
30
25
20
15
10
050
Pareto frontierProposed algorithm
02 0301 04 05 07 0906 08 10
f2
f1
(a) CTP1 calculated by the proposed algorithm
CNSGA-IIPareto frontier
40
35
30
25
20
15
10
05
f2
f1
0 02 0301 04 05 07 0906 08 10
(b) CTP1 calculated by CNSGA-II
1211100908070605040302
f2
Pareto frontierProposed algorithm
f1
0 02 0301 04 05 07 0906 08 10
(c) CTP2 calculated by the proposed algorithm
CNSGA-IIPareto frontier
1211100908070605040302
f2
f1
0 02 0301 04 05 07 0906 08 10
(d) CTP2 calculated by CNSGA-II
50
40
30
20
10
00 50 100 150
Pareto frontierProposed algorithm
f2
f1
(e) BNH calculated by the proposed algorithm
Pareto frontier
50
40
30
20
10
00 50 100 150
f2
f1
CNSGA-II
(f) BNH calculated by CNSGA-II
Figure 6 Continued
10 Journal of Applied Mathematics
15
10
05
00 05 10 15
Pareto frontierProposed algorithm
f2
f1
(g) TNK calculated by the proposed algorithm
Pareto frontier
15
10
05
00 05 10 15
f2
f1
CNSGA-II
(h) TNK calculated by CNSGA-II
Figure 6 The comparison chart of Pareto frontier
Table 4 The comparison of performance
Test function Algorithm 120574 SP
CTP1 CNSGA-Π 0021317 plusmn 0000323 0873321 plusmn 008725
The proposed 0009836 plusmn 0000410 0567933 plusmn 001845
CTP2 CNSGA-Π 0011120 plusmn 0000753 078314 plusmn 002843
The proposed 0007013 plusmn 0000554 0296543 plusmn 000453
BNH CNSGA-Π 0014947 plusmn 0000632 0336941 plusmn 000917
The proposed 0013766 plusmn 0000043 0209321 plusmn 000561
TNK CNSGA-Π 0013235 plusmn 0000740 0464542 plusmn 000730
The proposed 0006435 plusmn 0000017 0224560 plusmn 000159
Feed water Water
Feed ore Ball mill OverflowMixing
PumpClassifierwater
additionFirst
classifierrecycle First-
overflow Spiral classifier 2Second-stage
to flotationSpiral classifier 1 Second classifier recycle(rough concentrate)
addition
pump
stageoverflow
Figure 7 Flow diagram of the grinding and classification process
algorithms based on the two criterions 120574 and SP as shownin Table 4 It can be observed from the data in Table 4that the proposed algorithm performs significantly betterthan the classical CNSGA-II algorithm in convergence anddistribution uniformityThe simulation results show that thisalgorithm can accurately converge to global Pareto solutionsand can maintain diversity of population
4 Optimization of Grinding andClassification Process
41 Bauxite Grinding and Classification Process Thegrindingand classification process is the key preparation for thebauxite mineral processing Here we consider a bauxitegrinding process in a certain mineral company with singlegrinding and two-stage classification as shown in Figure 7
The process consists in a grinding ball mill and twospiral classifiers First classifier recycle will be put back tothe ball mill for regrinding and the first-stage overflowwill be put into second spiral classifier after being mixedwith water the second classifier recycle will be prepared forBayer production as the rough concentrate and the second-stage overflow will be sent to the next flotation processThe production objectives are composed of the productionyields technically represented by the solid flow of feed oresince the process is nonstorable and the mineral processingquality represented by percentage of the small-size fractionsof mineral particles in the second-stage overflows
Journal of Applied Mathematics 11
Table 5 Notations for the model of the grinding and classificationprocess
Notation Description
119901119894
Particle percentage of the 119894th size fraction inthe ball mill overflow
119891119895
Particle percentage of the 119894th size fraction inthe Feed ore
119862 Rate of the first classifier recycle119864119886119894 The efficiency of the first spiral classifier
120591 Themean residence time1198751 The internal concentration in ball mill119872MF The solid flow of feed ore1198821 The water addition of the first classifier recycle
1198822 The classifier water addition
119887119894119895 The breakage distribution function119878119894 The breakage rate function119889119894 The particle with the 119894th size
120572119888119894
The particle percentage of the 119894th size in thefirst classifier overflow
119889119900 The unit size of the particle1198752 First-stage overflow120601119861 Ball filling rate
119860119888 119861119888
Parameters of the first-stage classifier overflowsize fraction distribution
119864119886119894 The efficiency of the first spiral classifier
11988950119888
The particle size fraction after correctionseparation
119898 The separation accuracy119896 The intermix index
1198641198861015840
min 1198891015840
50119888 1198981015840 1198961015840 The corresponding key parameters to theefficiency of the second spiral classifier
119886 120572 120583 ΛFour key parameters to control the breakagerate function
119889min 119889max Theminimum and maximum particle sizes119860MF 119861MF Parameters of feed ore size fraction distribution
119865(119894)The cumulative particle percentage less thanthe 119894th size fraction in feed ore
1205721198881015840
119894
The particle percentage of the 119894th size fractionin the second classifier overflow
1198641198861015840
119894 The efficiency of the second spiral classifier
42 Predictive Model of the Grinding and Classification Pro-cess Here we establish the mathematical predictive modelof each unit process in the bauxite grinding and classificationprocess The notations of the indexes decision variables andparameters are listed in Table 5 These notations will be usedfor the model of the grinding and classification process
421 Ball Mill Circuit Model Here 119901119894is the particle per-
centage of 119894th size fraction in the ball mill overflow 119891119895is the
particle percentage of 119894th size fraction in feed ore rate of thefirst classifier recycle 119862 is known and 119864119886
119894is the efficiency of
the first spiral classifier According to a technical report offield investigation and study we have that
119901119894(1 + 119862) =
119889119894119894119891119894+ sum119894minus1
119895=1119894gt1119889119894119895[119864119886119894119901119895(1 + 119862) + 119891
119895]
1 minus 119889119894119894119864119886119894
119889119894119895=
119890119895 119894 = 119895
119894minus1
sum
119896=119895
119888119894119896119888119895119896(119890119896minus 119890119894) 119894 gt 119895
119888119894119895=
minus
119895minus1
sum
119896=119894
119888119894119896119888119895119896 119894 lt 119895
1 119894 = 119895
1
119878119894minus 119878119895
119894minus1
sum
119896=119895
119878119896119887119894119896119888119896119895 119894 gt 119895
119890119895=
1
(1 + 05 sdot 120591119878119895) (1 + 025 sdot 120591119878
119895)2 120591 =
81198751
119872MF (12)
where 120591 is the mean residence time of minerals 1198751is the
internal concentration in ball mill and
1198751=
119872MF (1 + 119862)
119872MF (1 + 119862) +1198821 + 03 (1198821 +1198822) (13)
where119872MF (119905ℎ) is the solid flow of feed ore1198821is the water
addition of the first classifier recycle and1198822is the classifier
water addition 119887119894119895is the breakage distribution function 119878
119894
is the breakage rate function and it satisfied the followingequation
119878119894=
(119886(119889119894119889119900)120572
)
(1 + (119889119894120583)Λ
)
(14)
where 119889119894is the particle with the 119894th size 119889
119900it is a unit when
per millimeter is a unit 119889119900= 1 119889
119894= 119894 (mm) and 119886 120572 120583
and Λ are four key parameters to control the breakage ratefunction
In a concrete grinding and classification process the ballmill size is fixed and the speed of ball mill is constantThrough data acquisition and testing of grinding and clas-sification steady-state loop the regression model between 119886120572 120583 Λ and condition variables size fraction distributioncan be established The input variables are ball filling rate120601119861 solid flow of feed ore 119872MF water addition of the first
classifier recycle1198821 and parameters of feed ore size fraction
distribution 119860MF 119861MF The regression model is
[119886 120572 120583 Λ]119879
=[[
[
11990911
sdot sdot sdot 1199091119895
d
1199091198941
sdot sdot sdot 119909119894119895
]]
]
sdot [1198821 120601119861 119860MF 119861MF 119872MF]119879
(15)
The value of 119909119894119895(119894 = 1 2 3 4 119895 = 1 2 5) can be
obtained by the experimental data regression 119860MF 119861MF can
12 Journal of Applied Mathematics
be obtained from feed ore size fraction distribution and 119865(119894)is the cumulative particle percentage less than the 119894th sizefraction in feed ore and it is represented as follows
119865 (119894) = 1 minus exp (minus119860MF119889119861MF119894
) (16)
422 Spiral Classifier Model 120572119888119894is the particle percentage
of the 119894th size fraction in the first classifier overflow and 119901119894
is the particle percentage of the 119894th size fraction in ball milloverflow The spiral classifier model is as follows
120572119888119894=
119901119894times (1 minus 119864119886
119894)
sum119894=1
(119901119894times (1 minus 119864119886
119894))times 100 (17)
where 119864119886119894is the efficiency of the first spiral classifier and the
mechanism formula of 119864119886119894is shown as follows
119864119886119894= 1 minus exp [minus0693(
119889119894minus 119889min
11988950119888
minus 119889min)
119898
]
+ 119864119886min sdot [1 minus (119889119894minus 119889min119889max
)]
119896
(18)
where 119889119894is the particle with the 119894th size 119889min and 119889max
represent maximum and minimum particle sizes 11988950119888
isthe particle size fraction after correction separation 119898 isseparation accuracy and 119896 is intermix index
Through data acquisition and testing of grinding andclassification steady-state loop the regressionmodel betweenclassification parameters and condition variables size frac-tion distribution can be established The input variablesinclude the solid flow of feed ore 119872MF the classifier wateraddition 119882
2 and the parameters of ball mill overflow size
fraction distribution 119860MF 119861MF The regression model isshown as follows
[119864119886min 11988950119888
119898 119896]119879
=[[
[
11991011
sdot sdot sdot 1199101119895
d
1199101198941
sdot sdot sdot 119910119894119895
]]
]
sdot [119872MF 1198822119860MP 119861MP]
119879
(19)
where the value of 119910119894119895(119894 = 1 2 3 4 119895 = 1 2 3 4) can be
obtained by data regressionThe first-stage overflow 119875
2calculation formula is as
follows
1198752=
119872MF(119872MF +1198821 +1198822)
(20)
Similarly we can get the second spiral classifier model asfollows
1205721198881015840
119894=
120572119888119894times (1 minus 119864119886
1015840
119894)
sum119894(120572119888119894times (1 minus 1198641198861015840))
times 100 (21)
where 1205721198881015840119894is the particle percentage of the 119894th size fraction in
the second classifier overflow 120572119888119894is the particle percentage
of the 119894th size fraction in the first classifier overflow and
1198641198861015840
119894is the efficiency of the second spiral classifier The spiral
classifier model is as follows
1198641198861015840
119894= 1 minus exp[
[
minus0693(1198891015840
119894minus 1198891015840
min1198891015840
50119888minus 1198891015840
min)
1198981015840
]
]
+ 1198641198861015840
min sdot [1 minus (1198891015840
119894minus 1198891015840
min1198891015840max
)]
1198961015840
(22)
where 1198641198861015840min 1198891015840
50119888 1198981015840 and 119896
1015840 are key parameters to theefficiency of the second spiral classifier Through data acqui-sition and testing of grinding and classification steady-stateloop the regression model between classification parametersand condition variables size fraction distribution can beestablished The input variables include solid flow of feedore119872MF and parameters of the first-stage classifier overflowsize fraction distribution 119860
119888 119861119888 which are solved by similar
equation to (20) The regression model is shown as follows
[1198641198861015840
min 1198891015840
5011988811989810158401198961015840]119879
=
[[[
[
1199101015840
11sdot sdot sdot 1199101015840
1119895
d
1199101015840
1198941sdot sdot sdot 119910
1015840
119894119895
]]]
]
sdot [119872MF 119860119888119861119888]119879
(23)
where the value of1199101015840119894119895(119894 = 1 2 3 4 119895 = 1 2 3) can be obtained
by experimental data regression
43 Optimization Model of Grinding and Classification Pro-cess Two objective functions in the process are identifiedone is tomaximize output119891
1(119883) and the other is tomaximize
the small-size fractions (less than 0075mm fractions) in thesecond-stage overflow 119891
2(119883) It is also necessary to ensure
that the grinding productmeets all of other technical require-ments and the least disturbance in the following flotationcircuit As the constraints the feed load of the grinding circuit119872MF the steel ball filling rate 120601
119861 the first and the second
overflows 1198751and 119875
2 and the particle percentage of fine size
fraction in the first and the second classifier overflows 120572119888minus0075
and 1205721198881015840minus0075
should be within the user specified boundsThe operation variables are the solid flow of feed ore
119872MF water addition of the first classifier recycle 1198821 ball
filling rate 120601119861 and water addition of the second classifier119882
2
Based on all of the above grinding and classification processmultiobjective optimization model is as follows
max 119865 = max [1198911(119883) 119891
2(119883)]
1198911(119883) = 119872MF
1198912(119883) = 120572119888
1015840
minus0075
= 119891 (119872MF11988211198822 120601119861)
st 119872MFmin le 119872MF le 119872MFmax120601119861min le 120601119861 le 120601119861max1198751min le 1198751 le 1198751max1198752min le 1198752 le 1198752max120572119888minus0075
ge 120572119888min
1205721198881015840
minus0075ge 1205721198881015840
min
(24)
Journal of Applied Mathematics 13
Table 6 The optimization results calculated by the proposed algo-rithm
Number 1198911(119883) (th) 119891
2(119883) ()
1 92972 908142 91810 919003 92360 909234 90620 934605 91310 924946 90170 938007 89390 952008 89480 949329 91530 9240010 89298 9576311 89824 9454912 89800 9439513 89710 9461814 91170 9280015 91880 9184016 92190 9128117 89870 9409018 91000 9328519 91960 9152020 91124 93221
85
87
89
91
93
95
97
74000 79000 84000 94000
Optimization resultsThe raw data
f1(x)
f2(x)
()
89000
Figure 8 The comparison chart between optimization results andindustrial data
In (24) 119891(119872MF11988211198822 120601119861) implicates the model ofgrinding and classification represented by (12)ndash(23) Theproposed algorithm is applied to solve the problem and theoptimization results are shown in Table 6
With the practical process data from a grinding circuitof a mineral plant the simulation of this hybrid intelligentmethod adopted the same parameters on the variation infresh slurry feed velocity density particle size distributionand cyclone feed operating configurations
The comparison of production data and optimizationresults in Table 6 is shown in Figure 8 where ldquoXrdquo repre-sents the proposed algorithm optimization results and ldquoIrdquorepresents the original data collected from the field withoutoptimization of raw data According to the objectives thedata point closer to the upper right edge is more beneficialObviously the proposed optimization result is far betterthan the original data indicating the effectiveness of theoptimization approach
44 TOPSIS Method for Solution Selection The resolution ofamultiobjective optimization problemdoes not endwhen thePareto-optimal set is found In practical operational prob-lems a single solution must be selected TOPSIS [36] is auseful technique in dealing with multiattribute or multicrite-ria decision-making (MADMMCDM) problems in the realworld The standard TOPSIS method attempts to choosealternatives that simultaneously have the shortest distancefrom the positive-ideal solution and the farthest distancefrom the negative-ideal solution According to the TOPSISmethod the relative closeness coefficient is calculated andthe best solution in Table 6 is the solution number 10 as1198911(119883) = 89298 (119905ℎ) and 119891
2(119883) = 95763 () The
corresponding decision variables are 119872MF = 89298 (119905ℎ)120601119861= 32119882
1= 75127 (119905ℎ) and119882
2= 16296 (119905ℎ)
5 Conclusions
Promoted by the requirements of engineering optimizationin complex practical processes of grinding and classificationwe proposed a hybrid multiobjective differential evolutionalgorithm with a few beneficial features integrated Firstly anarchiving mechanism for infeasible solutions is establishedwith functional partitioned multi-population which aims todirect the population to approach or land in the feasibleregion from different directions during evolution Secondlywe propose an infeasible constraint violations function toselect infeasible population with better performance sothat they are allowed to be saved and to participate inthe subsequent evolution Thirdly a nondominating rankingstrategy is designed to improve the crowding-distance sortingand return uniform distribution of Pareto solutions Finallythe simplex method is inserted in the differential evolutionprocess to purposefully enrich the diversity without excessivecomputation cost The advantage of the proposed algorithmis the exemption from constructing penalty function andthe preservation of meaningful infeasible solutions directlySimulation results on benchmarks indicate that the proposedalgorithm can converge quickly and effectively to the truePareto frontier with better distribution
Based on the investigated information about grinding cir-cuit process we established a multiobjective optimal modelwith equations frommechanism knowledge parameters rec-ognized by data regression and constraints of technicalrequirements The nonlinear multiobjective optimizationmodel is too complicated to be solved by traditional gradient-based algorithmsThe proposed hybrid differential evolutionalgorithm is applied and tested to achieve a Pareto solution
14 Journal of Applied Mathematics
set It is proven to be valuable for operation decision makingin the industrial process and showed superiority to the oper-ation carried out in the production In fact many operatingparameters in complex processes are highly coupled andconflicting with each other The optimal operation of theentire production process is very difficult to obtain bymanualcalculation let alone the fluctuation situation of processconditions The application case indicates that the proposedmethod has good performance and is helpful to inspirefurther research on evolutionary methods for engineeringoptimization
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgments
This work is supported by the National Natural ScienceFoundation of China (61374156 61273187 and 61134006)the National Key Technology Research and DevelopmentProgram of the Ministry of Science and Technology ofChina (2012BAK09B04) the Science Fund for CreativeResearch Groups of the National Natural Science Foundationof China (61321003) the Fund for Doctor Station of theMinistry of Education (20110162130011 and 20100162120019)and theHunanProvince Science andTechnology PlanProject(2012CK4018)
References
[1] K Mitra and R Gopinath ldquoMultiobjective optimization ofan industrial grinding operation using elitist nondominatedsorting genetic algorithmrdquo Chemical Engineering Science vol59 no 2 pp 385ndash396 2004
[2] G Yu T Y Chai and X C Luo ldquoMultiobjective productionplanning optimization using hybrid evolutionary algorithmsfor mineral processingrdquo IEEE Transactions on EvolutionaryComputation vol 15 no 4 pp 487ndash514 2011
[3] T Y Ma W H Gui Y L Wang and C H Yang ldquoDynamicoptimization control for grinding and classification processrdquoControl andDecision vol 27 no 2 pp 286ndash290 2012 (Chinese)
[4] F Y Edgeworth Mathematical Physics An Essay on the Appli-cation of Mathematics to the Moral Sciences C Kegan Paul ampCompany London UK 1881
[5] V Pareto Cours drsquoEconomie Politique F Rouge LausanneSwitzerland 1896
[6] X F Chen W H Gui Y L Wang and L Cen ldquoMulti-stepoptimal control of complex process a genetic programmingstrategy and its applicationrdquo Engineering Applications of Artifi-cial Intelligence vol 17 no 5 pp 491ndash500 2004
[7] C A C Coello ldquoEvolutionary multi-objective optimizationa historical view of the fieldrdquo IEEE Computational IntelligenceMagazine vol 1 no 1 pp 28ndash36 2006
[8] K Deb A Pratap S Agarwal and T Meyarivan ldquoA fastand elitist multiobjective genetic algorithm NSGA-IIrdquo IEEETransactions on Evolutionary Computation vol 6 no 2 pp 182ndash197 2002
[9] C M Fonesca and P J Fleming ldquoGenetic algorithms formulti objective optimization formulation discussion and gen-eralizationrdquo in Proceedings of the 5th International Conferenceon Genetic Algorithms pp 415ndash423 Morgan Kaufmann SanMateo Calif USA 1993
[10] S Sengupta S Das M Nasir A V Vasilakos and W PedryczldquoEnergy-efficient differentiated coverage of dynamic objectsusing an improved evolutionary multi-objective optimizationalgorithm with fuzzy-dominancerdquo in Proceedings of IEEECongress on Evolutionary Computation pp 1ndash8 IEEE PressBrisbane Australia 2012
[11] K V Price R M Storn and J A Lampinen DifferentialEvolution a Practical Approach to Global Optimization SpringerBerlin Germany 2005
[12] W Y Gong and Z H Cai ldquoAn improved multiobjective dif-ferential evolution based on Pareto-adaptive 120576-dominance andorthogonal designrdquo European Journal of Operational Researchvol 198 no 2 pp 576ndash601 2009
[13] X H Zeng W K Wong and S Y Leung ldquoAn operatorallocation optimization model for balancing control of thehybrid assembly lines using Pareto utility discrete differentialevolution algorithmrdquo Computers and Operations Research vol39 no 5 pp 1145ndash1159 2012
[14] E Zitzler and L Thiele ldquoMultiobjective evolutionary algo-rithms a comparative case study and the strength Paretoapproachrdquo IEEETransactions onEvolutionaryComputation vol3 no 4 pp 257ndash271 1999
[15] A Al-Ani A Alsukker and R N Khushaba ldquoFeature subsetselection using differential evolution and a wheel based searchstrategyrdquo Swarm and Evolutionary Computation vol 9 pp 15ndash26 2013
[16] G Y Li and M G Liu ldquoThe summary of differential evolutionalgorithm and its improvementsrdquo in Proceedings of the 3rdInternational Conference on Advanced Computer Theory andEngineering (ICACTE rsquo10) pp V3153ndashV3156 Chengdu ChinaAugust 2010
[17] H A Abbass ldquoThe self-adaptive Pareto differential evolutionalgorithmrdquo in Proceedings of the IEEE Congress on EvolutionaryComputation pp 831ndash836 IEEE Press Honolulu Hawaii USAMay 2002
[18] N K Madavan ldquoMultiobjective optimization using a Paretodifferential evolution approachrdquo in Proceedings of the IEEECongress on Evolutionary Computation pp 1145ndash1150 IEEEPress Honolulu Hawaii USA May 2002
[19] F Xue A C Sanderson and R J Graves ldquoPareto-based multi-objective differential evolutionrdquo in Proceedings of the IEEECongress on Evolutionary Computation pp 862ndash869 IEEEPress Piscataway NJ USA December 2003
[20] T Robic and B Filipi ldquoDEMO differential evolution formultiobjective optimizationrdquo in Evolutionary Multi-CriterionOptimization pp 520ndash533 Springer Berlin Germany 2005
[21] V L Huang A K Qin P N Suganthan and M F TasgetirenldquoMulti-objective optimization based on self-adaptive differen-tial evolution algorithmrdquo in Proceedings of the IEEE Congresson Evolutionary Computation (CEC rsquo07) pp 3601ndash3608 IEEEPress Singapore September 2007
[22] V L Huang S Z Zhao R Mallipeddi and P N SuganthanldquoMulti-objective optimization using self-adaptive differentialevolution algorithmrdquo in Proceedings of the IEEE Congress onEvolutionary Computation (CEC rsquo09) pp 190ndash194 IEEE PressTrondheim Norway May 2009
Journal of Applied Mathematics 15
[23] J A Adeyemo and F A O Otieno ldquoMulti-objective differentialevolution algorithm for solving engineering problemsrdquo Journalof Applied Sciences vol 9 no 20 pp 3652ndash3661 2009
[24] Y Wang and Z Cai ldquoCombining multiobjective optimizationwith differential evolution to solve constrained optimizationproblemsrdquo IEEE Transactions on Evolutionary Computationvol 16 no 1 pp 117ndash134 2012
[25] A Homaifar C X Qi and S H Lai ldquoConstrained optimizationvia genetic algorithmsrdquo Simulation vol 62 no 4 pp 242ndash2531994
[26] J A Joines and C R Houck ldquoOn the use of non-stationarypenalty functions to solve nonlinear constrained optimizationproblems with GArsquosrdquo in Proceedings of the 1st IEEE Conferenceon Evolutionary Computation pp 579ndash584 Orlando Fla USAJune 1994
[27] F Jimenez and J L Verdegay ldquoEvolutionary techniques forconstrained optimization problemsrdquo in Proceedings of the 7thEuropean Congress on Intelligent Techniques and Soft Computing(EUFIT rsquo99) Aachen Germany 1999
[28] C Sun J Zeng and J Pan ldquoAn improved vector particleswarm optimization for constrained optimization problemsrdquoInformation Sciences vol 181 no 6 pp 1153ndash1163 2011
[29] Z Michalewicz and M Schoenauer ldquoEvolutionary algorithmsfor constrained parameter optimization problemsrdquo Evolution-ary Computation vol 4 no 1 pp 1ndash32 1996
[30] C A C Coello and A D Christiansen ldquoMoses a multiob-jective optimization tool for engineering designrdquo EngineeringOptimization vol 31 no 1ndash3 pp 337ndash368 1999
[31] Y G Woldesenbet G G Yen and B G Tessema ldquoConstrainthandling in multiobjective evolutionary optimizationrdquo IEEETransactions on Evolutionary Computation vol 13 no 3 pp514ndash525 2009
[32] J A Nelder and R Mead ldquoA simplex method for functionminimizationrdquoThe Computer Journal vol 7 no 4 pp 308ndash3131965
[33] D A van Veldhuizen and G B Lamont ldquoMultiobjective evolu-tionary algorithms analyzing the state-of-the-artrdquo EvolutionaryComputation vol 8 no 2 pp 125ndash147 2000
[34] R Farmani D A Savic andG AWalters ldquoEvolutionarymulti-objective optimization in water distribution network designrdquoEngineering Optimization vol 37 no 2 pp 167ndash183 2005
[35] K Deb and T Goel Controlled Elitist Non-Dominated SortingGenetic Algorithms for Better Convergence Springer BerlinGermany 2001
[36] Y J Lai T Y Liu and C L Hwang ldquoTopsis for MODMrdquoEuropean Journal of Operational Research vol 76 no 3 pp486ndash500 1994
Submit your manuscripts athttpwwwhindawicom
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical Problems in Engineering
Hindawi Publishing Corporationhttpwwwhindawicom
Differential EquationsInternational Journal of
Volume 2014
Applied MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Probability and StatisticsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical PhysicsAdvances in
Complex AnalysisJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
OptimizationJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
CombinatoricsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Operations ResearchAdvances in
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Function Spaces
Abstract and Applied AnalysisHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of Mathematics and Mathematical Sciences
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Algebra
Discrete Dynamics in Nature and Society
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Decision SciencesAdvances in
Discrete MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014 Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Stochastic AnalysisInternational Journal of
2 Journal of Applied Mathematics
Multiobjective optimization problems (MOPs) in thecase of traditional optimization methods are often handledby aggregating multiple objectives into a single scalar objec-tive through weighting factors MOPs have a set of equallygood (nondominating) solutions instead of a single onecalled a Pareto optimum which was introduced by Edge-worth in 1881 [4] and later generalized by Pareto in 1896[5] The practical MOPs are often implicated in series ofequations functions or procedures with complicated con-straints Therefore the evolutionary algorithms are attractiveapproaches for low requirements onmathematical expression[6] Since the mid 1980s there has been a growing interestin solving MOPs using evolutionary approaches [7ndash10]One of the most successful evolutionary algorithms for theoptimization of continuous space functions is the differentialevolution (DE) [11] DE is simple and efficiently converges tothe global optimum in most cases [12 13] Its efficiency hasbeen proven [14] in many application fields such as patternrecognition [15] and mechanical engineering [16]
There have been many improvements for DE to solveMOPs Abbass [17] firstly provided a Pareto DE (PDE) algo-rithm for MOPs in which DE was employed to create newsolutions and only the nondominated solutions were kept asthe basis for the next generation Madavan [18] developed aPareto differential evolution approach (PDEA) in which newsolutions were created by DE and kept in an auxiliary popu-lation Xue et al [19] introduced multiobjective differentialevolution (MODE) and used Pareto-based ranking assign-ment and crowding distancemetric but in a differentmannerfrom PDEA Robic and Filipi [20] also adopting Pareto-based ranking assignment and crowding distance metricdeveloped a DE for multiobjective optimization (DEMO)with a different population update strategy and achievedgood results Huang et al [21] extended the self-adaptive DE(SADE) to solve MOPs by a so called multiobjective self-adaptive DE (MOSADE) They further extended MOSADEby using objectivewise learning strategies [22] Adeyemo andOtieno [23] provided multiobjective differential evolutionalgorithm (MDEA) In MDEA a new solution was generatedby DE variant and compared with target solution If itdominates the target solution then it was added to the newpopulation otherwise a target solution was added
On the other hand single-objective constrained opti-mization problems have been studied intensively in the pastyears [24ndash28] Different constraint handling techniques havebeen proposed to solve constrained optimization problemsMichalewicz and Schoenauer [29] divided constraints han-dling methods used in evolutionary algorithms into fourcategories preserving feasibility of solutions penalty func-tions separating the feasible and infeasible solutions andhybrid methods The differences among these methods arehow to deal with the infeasible individuals throughout thesearch phases Currently the penalty functionmethod ismostwidely used and this algorithm strongly depends on thechoice of the penalty parameter
Although the multiobjective optimization and the con-straint handling problem have received lots of contributionrespectively the CMOPs are still difficult to be solved inpractice Coello and Christiansen [30] proposed a simple
approach to solve CMOPs by ignoring any solution that vio-lates any of the assigned constraints Deb et al [8] proposeda constrained multiobjective algorithm based on the conceptof constrained domination which is also known as superi-ority of the feasible solution Woldesenbet et al [31] intro-duced a constraint handling technique based on adaptivepenalty functions and distance measures by extending thecorresponding version for the single-objective constrainedoptimization
In the MOP of grinding and classification process thedefinitions of Pareto solutions Pareto frontier and Paretodominance are in consistency with the classic definitionsClearly the Pareto frontier is amapping of the Pareto-optimalsolutions to the objective space In the minimization sensegeneral constrained MOPs can be formulated as follows
min 119865 (119883) = min [1198911(119883) 119891
2(119883) 119891
119903(119883)]
st 119892119894(119883) le 0 (119894 = 1 2 119901)
ℎ119895(119883) = 0 (119895 = 119901 + 1 119902)
119909119896isin [119909119896min 119909119896max] (119896 = 1 2 119899)
(1)
where 119865(119883) is the objective vector 119883 = (1199091 119909
119899) isin R119899 is
a parameter vector 119892119894(119883) is the 119894th inequality constraint
and ℎ119895(119883) is the 119895th equality constraint 119909
119896min and 119909119896max
are respectively the lower and upper bounds of the decisionvariable 119909
119896
In this paper based on the specific industrial backgroundof continuous bauxite grinding-classification operation anew hybrid DE algorithm is proposed to solve complex con-strained multiobjective optimization problems Firstly ahybrid DE algorithm for MOPs with simplex method (SM-DEMO) is designed to overcome the problems of globalperformance degradation and being trapped in local opti-mum Then for the MOPs with complicated constraints theproposed algorithm is formed by combining SM-DEMO andfunctional partitioned multi-population In this method theconstruction of penalty functions is not required and themeaningful infeasible solutions are fully utilized
The remainder of the paper is structured as followsSection 2 describes the SM-DEMO algorithm for uncon-strained cases The proposed algorithm of multipopulationfor constrained MOPs is given in Section 3 with verificationof performance by benchmark testing results Section 4describes the model of productsrsquo output and quality in thegrinding-classification process in detail and the applicationof the proposed algorithm in the optimizationmodel Finallythe conclusions based on the present study are drawn inSection 5
2 SM-DEMO Algorithm forUnconstrained MOPs
In order to efficiently solve multiobjective optimizationproblem and find the approximately complete and nearop-timal Pareto frontier we proposed a hybrid DE algorithmfor unconstrained multiobjective optimization with simplexmethod
The differential evolution with initialization crossoverand selection as in usual genetic algorithms uses a pertur-bation of two members as the mutation operator to produce
Journal of Applied Mathematics 3
a new individualThemutation operator of the DE algorithmis described as follows
Considering each target individual 119909119866119894 in the 119866th gener-
ation of size119873119901 a mutant individual 119909119866+1119894
is defined by
119909119866+1
119894= 119909119866
1199033+ 119865 (119909
119866
1199031minus 119909119866
1199032) (2)
where indexes 1199031 1199032 and 119903
3represent mutually different
integers that are different from 119894 and that are randomlygenerated over [1119873119901] and 119865 is the scaling factor
The simplex method proposed by Spendley Hext andHimsworth and later refined by Nelder andMead (NM) [32]is a derivative-free line-search method that is particularlydesigned for traditional unconstrained minimization scenar-ios Clearly NM method can be deemed as a direct line-search method of the steepest descent kind The ingredientsof the replacement process consist of four basic operationsreflection expansion contraction and shrinkage Throughthese operations the simplex can improve itself and approxi-mate to a local optimum point sequentially Furthermore thesimplex can vary its shape size and orientation to adapt itselfto the local contour of the objective function
21 Main Strategy of SM-DEMO The SM-DEMO algorithmis improved by the following three points compared with DE
211 Modified Selection Operation After traditional DE evo-lution the individual 119906119866+1
119894119895may violate the boundary con-
straints 119909max119894119895
and 119909min119894119895
119906119866+1119894119895
is replaced by new individual119908119866+1
119894119895being adjusted as follows
119908119866+1
119894119895
=
119909max119894119895
+ rand () lowast (119909max119894119895
minus 119906119866+1
119894119895) if (119906119866+1
119894119895gt 119909
max119894119895
)
119909min119894119895
+ rand () lowast (119909min119894119895
minus 119906119866+1
119894119895) if (119906119866+1
119894119895lt 119909
min119894119895
)
119906119866+1
119894119895 otherwise
(3)
The new population is combined with the existing parentpopulation to form a new set 119872119892 of bigger size than 119873119901A nondominated ranking of 119872119892 is performed and the 119873119901best individuals are selected This approach allows a globalnondomination checking between both the parent and thenew generation rather than only in the new generation asis done in other approaches whereas it requires additionalcomputational cost in sorting the combined
212 Nondominated Ranking Based on Euclidean DistanceThe solutions within each nondominated frontier that residein the less crowded region in the frontier are assigned a higherrank as the NSGA-II algorithm [8] developed by Deb et alindicated The crowding distance of the 119894th solution in itsfrontier (marked with solid circles) is the average side lengthof the cuboids (shown with a dashed box in Figure 1(a))The crowding-distance computation requires sorting thepopulation according to each objective function value inascending order of magnitude As shown in Figure 1 119860 and119862 are two solutions near 119861 in the same rank and 120590
0(119861) is the
crowding distance of the119861th solution traditionally calculatedas follows
1205900(119861) =
119899
sum
119895=1
10038161003816100381610038161003816119891119895(119860) minus 119891
119895(119862)
10038161003816100381610038161003816 (4)
where 119891119895(119860) 119891
119895(119862) are the objective vectors For each
objective function the boundary solutions (solutions withthe smallest and the largest function values) are assigned aninfinite distance value
A crowding-distance metric is used to estimate thedensity of solutions surrounding a particular solution in thepopulation and is obtained from the average distance of thetwo solutions on either side of the solution along each of theobjectives As shown in Figure 1119860119861119862 are the individuals ofthe generation on the same frontier and we easily know thatthe density in Figure 1(a) is better than that in Figure 1(b) Ifwe use (4) to calculate the crowding distance of 119862 we onlyknow that in Figure 1(a) it is better than in Figure 1(b) thecrowding distance of119862 in Figures 1(a) and 1(c) is equal whichis against the knowledge
To distinguish the mentioned situations we proposean improved crowding-distance metric based on Euclideandistance 119872 is the center point of the line 119860119861 119891
119895is the 119895th
objective vector and the crowding distance 120590(119861) is defined asfollows
120590 (119861) = |119860119862| minus |119861119872|
= radic
119899
sum
119895=1
[119891119895(119860) minus 119891
119895(119862)]2
minus radic
119899
sum
119895=1
119891119895(119861) minus
[119891119895(119860) + 119891
119895(119862)]
2
2
(5)
The crowded-comparison operator guides the selectionprocess at the various stages of the algorithm toward auniformly distributed Pareto-optimal frontier To carry outthe comparison we assume that every individual in thepopulation has two attributes (1) nondomination rank 119894rankand (2) crowding distance 119894distance Then a partial order isdefined as 119894 ≺ 119895 If 119894 ≺ 119895 that is between two solutionswith different nondomination ranks we prefer the solutionwith the lower (better) rank namely 119894rank ≺ 119895rank Otherwiseif both solutions belong to the same frontier that is 119894rank =
119895rank then we prefer the solution that is located in a lessercrowded region that is 119894distance ≻ 119895distance
213 Simplex Method for Local Search The simplex methodfor local search is mixed in the evolution process to enhancethe searching strategy in the optimization process The goalof integrating NM simplex method [32] and DE is to enrichthe population diversity and avoid being trapped in localminimum We apply NM simplex method operator to thepresent population when the number of iterations is greaterthan a particular value (like 119866max2) The individuals thatachieved the single extreme value in each objective functionaremarked as the initial vertex points of simplexmethodThe
4 Journal of Applied Mathematics
A
B
C
f1
f2
(a)
A
B
C
f1
f2
(b)
A
B
C
f1
f2
(c)
A
B
C
M
f1
f2
(d)
Figure 1 Crowding-distance diagram
present population is updated according to simplex methoduntil the terminal conditions are satisfied
The computation steps of the algorithm are included inSection 32
22 Evaluation Criteria Unlike the single-objective opti-mization it is more complicated for solution quality eval-uation in the case of multiobjective optimization Many ofthe suggested methods can be summarized in two typesOne is to evaluate the convergence degree by computing theproximity between the solution frontier and the actual Paretofrontier The other is to evaluate the distribution degree ofthe solutions in objective space by computing the distancesamong the individuals Here we choose both methods toevaluate the performance of the SM-DEMO algorithm
(1) Convergence Evaluation Deb et al [8] proposed thismethod in 2002 It is described as follows
120574 =1
119876(
119876
sum
119894=1
min 1003817100381710038171003817119875lowastminus 119875FT
1003817100381710038171003817) (6)
where 120574 is the extent of convergence to a known of Pareto-optimal set 119875lowast is the obtained nondomination Pareto fron-tier 119875FT is the real nondomination Pareto frontier 119875lowast minus119875FTis the Euclidean distance of 119875lowast with 119875FT and119876 is the numberof obtained solutions
(2) Distribution Degree EvaluationThenonuniformity in thedistribution is measured by SP as follows
SP = radic 1
(119876 minus 1)
119876
sum
119894=1
(119889 minus 119889119894)2
(7)
where 119889119894is the Euclidean distance among consecutive solu-
tions in the obtained nondominated set of solutions andparameter 119889 is the average distance
23 Experimental Studies Four well-known benchmark testfunctions [33] are used here to compare the performance ofSM-DEMO with NSGA-II DEMOParent These four prob-lems are called ZDT2 ZDT3 ZDT4 and ZDT6 each has twoobjective functions We describe them in Table 1
Journal of Applied Mathematics 5
Table 1 Test problems
Test problems Objective functionsmin119865(119883) = min[119891
1(119883) 119891
2(119883)]
Range of variable
ZDT21198911(119883) = 119909
1 1198912(119883) = 119892(119883)(1 minus (
1198911
119892(119883))
2
)
119892 (119883) = 1 + 9
119899
sum
119894=2
119909119894
119899 minus 1
119899 = 30
0 le 119909119894le 1
ZDT31198911(119883) = 119909
1 1198912(119883) = 119892(1 minus radic(
1198911
119892) minus (
1198911
119892) sin (10120587119891
1))
119892 (119883) = 1 + 9
119899
sum
119894=2
119909119894
119899 minus 1
119899 = 30
0 le 119909119894le 1
ZDT41198911(119883) = 119909
1 1198912(119883) = 119892(1 minus radic(
1198911
119892))
119892 (119883) = 1 + 10 (119899 minus 1) +
119899
sum
119894=2
(1199092
119894minus 10 cos (4120587119909
119894))
119899 = 10
0 le 1199091le 1
minus5 le 119909119894le 5 119894 = 2 119899
ZDT61198911(119883) = 1 minus exp (minus4119909
1) sin6 (6120587119909
1) 1198912(119883) = 119892(1 minus (
1198911
119892)
2
)
119892 (119883) = 1 + 9
119899
sum
119894=2
119909119894
(119899 minus 1)025
119899 = 10
0 le 119909119894le 1
Table 2 The performance results of the each algorithm on the test function
Test function Algorithm 120574 SP
ZDT2DEMOParent 0005120 plusmn 0000312 0000630 plusmn 0000010
NSGA-2 0007120 plusmn 0000413 0000540 plusmn 0000940
SM-DEMO 0004013 plusmn 0000230 0000423 plusmn 0000011
ZDT3DEMOParent 0009704 plusmn 0000027 0007512 plusmn 0000165
NSGA-2 0014067 plusmn 0000059 0006540 plusmn 0000124
SM-DEMO 0004704 plusmn 0000003 0004450 plusmn 0000153
ZDT4DEMOParent 2009704 plusmn 0901164 0011031 plusmn 0001104
NSGA-2 3144067 plusmn 2100740 0010122 plusmn 0000072
SM-DEMO 0874001 plusmn 0014323 0008721 plusmn 0000159
ZDT6DEMOParent 0649704 plusmn 0004912 0104520 plusmn 0015486
NSGA-2 1014067 plusmn 0010421 0007942 plusmn 0000105
SM-DEMO 0007750 plusmn 0000083 0002014 plusmn 0000117
The simulation is carried out under the environment ofIntel Pentium 4 CPU 306GHz 512MB memory WindowsXP Professional Matlab71 Initialization parameters are setas follows population size119873119901 = 100 scaling factor 119865 = 08cross rate 119862
119877= 06 maximum evolution generation 119866max =
250 and number of SM evolution iterations 119866SM = 100All of the three algorithms are real coded with equal
population size and equal maximum evolution generationEach algorithm independently runs 20 times for each testfunction Because we cannot get the real Pareto-optimal setwewill take 60 Pareto-optimal solutions obtained by the threealgorithms as a true Pareto-optimal solution set
We evaluated the algorithms based on the two perfor-mance indexes 120574 and SP Table 2 shows themean and varianceof 120574 and SP using three algorithms SM-DEMO NSGA-II and DEMOParent We can learn from Table 2 for theZDT2 function that all of the three algorithms have a good
performance while the SM-DEMO is slightly better than theother two algorithms In terms of convergences for ZDT3ZDT4 and ZDT6 which are more complex than ZDT2 SM-DEMO is significantly better thanDEMOParent andNSGA-II
Figure 2 shows a random running of SM-DEMO algo-rithm It is clear that SM-DEMO algorithm can produce agood approximation and a uniform distribution
3 Proposed Hybrid Algorithm for CMOP
The space of constrained multiobjective optimization prob-lem can be divided into the feasible solution space and theinfeasible solution space as shown in Figure 3 where 119878 isthe search space Ω is the feasible solution space and 119885 isthe infeasible solution space 119909
119894(119894 = 1 2 3 4) is the feasible
solution and 119910119894(119894 = 1 2 3 4) is the infeasible solution
6 Journal of Applied Mathematics
SM-DEMOPareto frontier
0 01 02 03 04 05 06 07 08 09 1f1
0
01
02
03
04
05
06
07
08
09
1f2
(a) ZDT2
0
02
04
06
08
1
minus08
minus06
minus04
minus02
f2
SM-DEMOPareto frontier
0 01 02 03 04 05 06 07 08 09 1f1
(b) ZDT3
SM-DEMOPareto frontier
0 01 02 03 04 05 06 07 08 09 1f1
0
01
02
03
04
05
06
07
08
09
1
f2
(c) ZDT4
SM-DEMO
0
01
02
03
04
05
06
07
08
09
1
Pareto frontier
02 03 04 05 06 07 08 09 1f1
f2
(d) ZDT6
Figure 2 SM-DEMO simulation curve
Assume that 119909lowast is the global optimal solution and 1199101is
the closest one to 119909lowast If the infeasible population 1199101is not
excluded by the evolution algorithm it is permitted to exploreboundary regions from new directions where the optimumis likely to be found
31 General Idea of the Proposed Algorithm Researchershave gradually realized the merit of infeasible solutions insearching for the global optimum in the feasible region Someinfeasible solutions with better performance are allowed to besaved Farmani et al [34] formulated a method to ensure thatinfeasible solutions with a slight violation become feasible inevolution Based on the constraints processing approach ofmultiobjective optimization problems the proposed hybridDE algorithm avoids constructing penalty function anddeleting meaningful infeasible solutions directly
Here the proposed algorithm will produce multiplegroups of functional partitions which include an evolution-ary population 119875119892 of size 119873119901 an intermediate population119872119892 to save feasible individuals an intermediate population119878119892 to save infeasible individuals a population 119875119891 to save theoptimal feasible solution found in the search process and apopulation 119875119888 to save the optimal infeasible solution Therelationship of multi-population is shown in Figure 4
With the description of (1) equality constraints are alwaystransformed into inequality constraints as |ℎ
119895(119883)| minus 120575 le 0
where 119895 = 119901 + 1 119902 and 120575 is a positive tolerance valueTo evaluate the infeasible solution the degree of constraintviolation of individual 119883 on the 119895th constraint is calculatedas follows
119881119894(119883) =
max 0 119892119894(119883) (119894 = 1 2 119901)
max 0 10038161003816100381610038161003816ℎ119895 (119883)10038161003816100381610038161003816minus 120575 (119895 = 119901 + 1 119902)
(8)
Journal of Applied Mathematics 7
Feasible solutionspace Ω
Search space S
x2
x1y1
y4
y3
y2 x3
x4
xlowast
Infeasible solutionspace Z
Figure 3 Distribution diagram of search space
Evolutionarypopulation
Pg
Intermediatepopulation
Mg
Optimal Randomly selected
Intermediatepopulation
SgInfeasible
Feasibleindividuals
individuals
Randomlyselected
Evolution
g = g + 1
g = g + 1
g = g + 1Randomly selected
feasiblesolution Pf
Optimal infeasible
solution Pc
Figure 4The relationship diagram ofmultipopulation demonstrat-ing
The final constraint violation of each individual in thepopulation can be obtained by calculating the mean of thenormalized constraint violations
In order to take advantage of the infeasible solutionswith better performance we proposed the following adaptivedifferential mutation operator to generate individual varia-tion learning from the mutation operators in DErand-to-best1bin according to rules defined by Price et al [11]Considering each individual vector 119909119866
119894 a mutant individual
119909119866+1
119894is defined by
119909119866+1
119894
= 119909119866
119894+ 1198651sdot (119909119866
1198911minus 119909119866
1199031)+ 1198652sdot (119909119866
1198912minus 119909119866
1199032) 119877
119862ge rand ( )
119909119866
119894+ 1198651sdot (119910119866
119894minus 119909119866
1199031)+ 1198652sdot (119909119866
1198911minus 119909119866
1199032) 119877
119862lt rand ( )
(9)
where 1199031and 1199032represent different integers and also different
from 119894 randomly generated over [1119873119901] 119865 is the scalingfactor 119909119866
119891119894(119894 = 1 2 119899) is randomly generated from 119875119891
119910119866
119894(119894 = 1 2 119899) is randomly generated from 119875119888 and 119877
119862is
the mutation factor as follows
119877119862= 1198771198880sdot120574119866+ const
120574119866minus1 + const (10)
where 1198771198880is the initial value of the variability factor const is
a small constant to ensure that the fractional is meaningfuland 120574119866 is defined as follows
120574119866=the number of infeasible solutions in 119875119892
119873119901 (11)
32 Framework of the Proposed Algorithm The proposedalgorithm is described as follows
Step 1 (initialization) Generate the population 119875119892 119875119891 and119875119888 of size 119873119901 119873119875
1 and 119873119875
2 Set the value of 119862
119877(crossover
probability)119866max (the number of function evaluations)119866SM(the iterative number of evolution by NM simplex method)119892 = 1 (the current generation number) and positive controlparameter for scaling the difference vectors 119865
1 1198652 Randomly
generate the parent population 119875119892 from the decision spaceSet the 119875119891 and 119875119888 and let the intermediate populations 119878119892and119872119892 be empty
Step 2 (DE reproduction) By (3) and (9) for mutationcrossover and selection an offspring 119878119892 is created Judge theconstraints of all individuals in 119875119892 In accordance with (8)we first calculate constraint violation degree 119881
119894(119883) of all of
the individuals If 119881119894(119883) = 0 the solution is feasible and
preserved to the intermediate set 119872119892 if 119881119894(119883) gt 0 the
solution is infeasible and preserved to the intermediate set 119878119892
Step 3 (simplex method) Apply NM simplex method oper-ator to the present population if 119892 ge 119866max2 Updatethe present population 119872119892 when the number of iterationsexceeds maximum iterations
Step 4 (119875119891 construction) Rank chromosomes in 119872119892 basedon (5) and generate the elitist population 119875119891 (the size is119873119901)from the ranked population119872119892
Step 5 (119875119888 construction) Add the chromosomes in 119878119892 withslight constraint violations to the 119875119888
Step 6 (mixing the population) Combine 119878119892with the existingparent population to form a new set119872119892 Remove the dupli-cate individuals in119872119892 and the existing parent population
Step 7 (evolution) Randomly choose chromosomes from119875119888 119875119892 and 119875119891 Use the adaptive differential mutation anduniform discrete crossover to obtain the offspring population119875119892 + 1
Step 8 (termination) If the stopping criterion ismet stop andoutput the best solution else go to Step 2
33 Experimental Study In this section we choose threeproblems CTP TNK and BNH as shown in Table 3 to testthe proposed method and compare the method with thecurrent CNSGA-II [35]
8 Journal of Applied Mathematics
Table 3 Test functions
Testfunction
Objective functionmin119865(119883) = min[119891
1(119883) 119891
2(119883)]
Constraints Range ofvariable
CTP
1198911(119883) = 119909
1
1198912(119883) = exp(minus
1198911(119883)
119888 (119883))
times41 +
5
sum
119894=2
[119909119894
2minus 10 cos (2120587119909
119894)]
1198921(119883) = cos (120579) (119891
2(119883) minus 119890) minus sin (120579) 119891
1(119883)
1198922(119883) = 119886
1003816100381610038161003816sin 119887120587 [sin (120579) (1198912 (119883) minus 119890)+ cos (120579) 119891
1(119883)]119888
10038161003816100381610038161003816
119889
1198921(119883) ge 119892
2(119883)
0 le 1199091le 1
minus5 le 119909119894le 5
119894 = 2 3 4 5
BNH1198911(119883) = 4119909
1
2+ 41199092
2
1198912(119883) = (119909
1minus 5)2+ (1199092minus 5)2
1198921(119883) = (119909
1minus 5)2+ 1199092
2minus 25
1198922(119883) = minus(119909
1minus 8)2+ (1199092+ 3) + 77
1198921(119883) le 0 119892
2(119883) le 0
0 le 1199091le 5
0 le 1199092le 3
TNK1198911(119883) = 119909
1
1198912(119883) = 119909
2
1198921(119883) = minus119909
1
2minus 1199092
2+ 1
+01 cos(16 arctan(11990911199092
))
1198922(119883) = (119909
1minus 05)
2+ (1199092minus 05)
1198921(119883) le 0 119892
2(119883) minus 05 le 0
0 le 119909119894le 120587
119894 = 1 2
20
15
10
5
00
Feasible solutionInfeasible solution
02 04 06 08 10f1
f2
Pareto-optimal frontier
(a) CTP1
20
15
10
05
0002 04 06 08 10
Feasible solutionInfeasible solution
Pareto-optimal frontier
f1
f2
(b) CTP2
Figure 5 CTP solution space
For CTP problem there are the six parameters 120579 119886 119887 119888119889 and 119890 that must be chosen in a way so that a portion ofthe unconstrained Pareto-optimal region is infeasible Eachconstraint is an implicit non-linear function of decisionvariables Thus it may be difficult to find a number ofsolutions on a non-linear constraint boundary We take twosets of values for six parameters in CTP problem which aredetermined as (1) CTP1 120579 = 01120587 119886 = 40 119887 = 05 119888 = 1119889 = 2 and 119890 = minus2 (2) CTP2 120579 = minus02120587 119886 = 02 119887 = 10119888 = 1 119889 = 6 and 119890 = 1 The Pareto frontiers the feasiblesolution spaces and the infeasible solution spaces are shownin Figure 5
The parameters are initialized as followsThe size of pop-ulation 119875119892 is119873119901 = 200 size of 119875119891 is119873119875
1= 150 size of 119875119888 is
1198731198752= 10 scaling factors 119865
1and 119865
2are randomly generated
within [05 1] cross rate is 119862119877= 06 maximum evolution
generation is 119866max = 300 and number of SM evolutioniterations is 119866SM = 100 All of the proposed algorithmsand CNSGA-II are real coded with equal population size andequal maximum evolution generation Each algorithm runs20 times independently for each test function
Figure 6 shows the result of a random running of theproposed algorithm and NSGA-II the smooth curve ldquomdashrdquorepresents the Pareto frontier and ldquoXrdquo stands for the solutionachieved by the proposed algorithm or NSGA-II
It is obvious that the proposed algorithm returns abetter approximation to the true Pareto-optimal frontie anda distribution of higher uniformity We also evaluated the
Journal of Applied Mathematics 9
40
35
30
25
20
15
10
050
Pareto frontierProposed algorithm
02 0301 04 05 07 0906 08 10
f2
f1
(a) CTP1 calculated by the proposed algorithm
CNSGA-IIPareto frontier
40
35
30
25
20
15
10
05
f2
f1
0 02 0301 04 05 07 0906 08 10
(b) CTP1 calculated by CNSGA-II
1211100908070605040302
f2
Pareto frontierProposed algorithm
f1
0 02 0301 04 05 07 0906 08 10
(c) CTP2 calculated by the proposed algorithm
CNSGA-IIPareto frontier
1211100908070605040302
f2
f1
0 02 0301 04 05 07 0906 08 10
(d) CTP2 calculated by CNSGA-II
50
40
30
20
10
00 50 100 150
Pareto frontierProposed algorithm
f2
f1
(e) BNH calculated by the proposed algorithm
Pareto frontier
50
40
30
20
10
00 50 100 150
f2
f1
CNSGA-II
(f) BNH calculated by CNSGA-II
Figure 6 Continued
10 Journal of Applied Mathematics
15
10
05
00 05 10 15
Pareto frontierProposed algorithm
f2
f1
(g) TNK calculated by the proposed algorithm
Pareto frontier
15
10
05
00 05 10 15
f2
f1
CNSGA-II
(h) TNK calculated by CNSGA-II
Figure 6 The comparison chart of Pareto frontier
Table 4 The comparison of performance
Test function Algorithm 120574 SP
CTP1 CNSGA-Π 0021317 plusmn 0000323 0873321 plusmn 008725
The proposed 0009836 plusmn 0000410 0567933 plusmn 001845
CTP2 CNSGA-Π 0011120 plusmn 0000753 078314 plusmn 002843
The proposed 0007013 plusmn 0000554 0296543 plusmn 000453
BNH CNSGA-Π 0014947 plusmn 0000632 0336941 plusmn 000917
The proposed 0013766 plusmn 0000043 0209321 plusmn 000561
TNK CNSGA-Π 0013235 plusmn 0000740 0464542 plusmn 000730
The proposed 0006435 plusmn 0000017 0224560 plusmn 000159
Feed water Water
Feed ore Ball mill OverflowMixing
PumpClassifierwater
additionFirst
classifierrecycle First-
overflow Spiral classifier 2Second-stage
to flotationSpiral classifier 1 Second classifier recycle(rough concentrate)
addition
pump
stageoverflow
Figure 7 Flow diagram of the grinding and classification process
algorithms based on the two criterions 120574 and SP as shownin Table 4 It can be observed from the data in Table 4that the proposed algorithm performs significantly betterthan the classical CNSGA-II algorithm in convergence anddistribution uniformityThe simulation results show that thisalgorithm can accurately converge to global Pareto solutionsand can maintain diversity of population
4 Optimization of Grinding andClassification Process
41 Bauxite Grinding and Classification Process Thegrindingand classification process is the key preparation for thebauxite mineral processing Here we consider a bauxitegrinding process in a certain mineral company with singlegrinding and two-stage classification as shown in Figure 7
The process consists in a grinding ball mill and twospiral classifiers First classifier recycle will be put back tothe ball mill for regrinding and the first-stage overflowwill be put into second spiral classifier after being mixedwith water the second classifier recycle will be prepared forBayer production as the rough concentrate and the second-stage overflow will be sent to the next flotation processThe production objectives are composed of the productionyields technically represented by the solid flow of feed oresince the process is nonstorable and the mineral processingquality represented by percentage of the small-size fractionsof mineral particles in the second-stage overflows
Journal of Applied Mathematics 11
Table 5 Notations for the model of the grinding and classificationprocess
Notation Description
119901119894
Particle percentage of the 119894th size fraction inthe ball mill overflow
119891119895
Particle percentage of the 119894th size fraction inthe Feed ore
119862 Rate of the first classifier recycle119864119886119894 The efficiency of the first spiral classifier
120591 Themean residence time1198751 The internal concentration in ball mill119872MF The solid flow of feed ore1198821 The water addition of the first classifier recycle
1198822 The classifier water addition
119887119894119895 The breakage distribution function119878119894 The breakage rate function119889119894 The particle with the 119894th size
120572119888119894
The particle percentage of the 119894th size in thefirst classifier overflow
119889119900 The unit size of the particle1198752 First-stage overflow120601119861 Ball filling rate
119860119888 119861119888
Parameters of the first-stage classifier overflowsize fraction distribution
119864119886119894 The efficiency of the first spiral classifier
11988950119888
The particle size fraction after correctionseparation
119898 The separation accuracy119896 The intermix index
1198641198861015840
min 1198891015840
50119888 1198981015840 1198961015840 The corresponding key parameters to theefficiency of the second spiral classifier
119886 120572 120583 ΛFour key parameters to control the breakagerate function
119889min 119889max Theminimum and maximum particle sizes119860MF 119861MF Parameters of feed ore size fraction distribution
119865(119894)The cumulative particle percentage less thanthe 119894th size fraction in feed ore
1205721198881015840
119894
The particle percentage of the 119894th size fractionin the second classifier overflow
1198641198861015840
119894 The efficiency of the second spiral classifier
42 Predictive Model of the Grinding and Classification Pro-cess Here we establish the mathematical predictive modelof each unit process in the bauxite grinding and classificationprocess The notations of the indexes decision variables andparameters are listed in Table 5 These notations will be usedfor the model of the grinding and classification process
421 Ball Mill Circuit Model Here 119901119894is the particle per-
centage of 119894th size fraction in the ball mill overflow 119891119895is the
particle percentage of 119894th size fraction in feed ore rate of thefirst classifier recycle 119862 is known and 119864119886
119894is the efficiency of
the first spiral classifier According to a technical report offield investigation and study we have that
119901119894(1 + 119862) =
119889119894119894119891119894+ sum119894minus1
119895=1119894gt1119889119894119895[119864119886119894119901119895(1 + 119862) + 119891
119895]
1 minus 119889119894119894119864119886119894
119889119894119895=
119890119895 119894 = 119895
119894minus1
sum
119896=119895
119888119894119896119888119895119896(119890119896minus 119890119894) 119894 gt 119895
119888119894119895=
minus
119895minus1
sum
119896=119894
119888119894119896119888119895119896 119894 lt 119895
1 119894 = 119895
1
119878119894minus 119878119895
119894minus1
sum
119896=119895
119878119896119887119894119896119888119896119895 119894 gt 119895
119890119895=
1
(1 + 05 sdot 120591119878119895) (1 + 025 sdot 120591119878
119895)2 120591 =
81198751
119872MF (12)
where 120591 is the mean residence time of minerals 1198751is the
internal concentration in ball mill and
1198751=
119872MF (1 + 119862)
119872MF (1 + 119862) +1198821 + 03 (1198821 +1198822) (13)
where119872MF (119905ℎ) is the solid flow of feed ore1198821is the water
addition of the first classifier recycle and1198822is the classifier
water addition 119887119894119895is the breakage distribution function 119878
119894
is the breakage rate function and it satisfied the followingequation
119878119894=
(119886(119889119894119889119900)120572
)
(1 + (119889119894120583)Λ
)
(14)
where 119889119894is the particle with the 119894th size 119889
119900it is a unit when
per millimeter is a unit 119889119900= 1 119889
119894= 119894 (mm) and 119886 120572 120583
and Λ are four key parameters to control the breakage ratefunction
In a concrete grinding and classification process the ballmill size is fixed and the speed of ball mill is constantThrough data acquisition and testing of grinding and clas-sification steady-state loop the regression model between 119886120572 120583 Λ and condition variables size fraction distributioncan be established The input variables are ball filling rate120601119861 solid flow of feed ore 119872MF water addition of the first
classifier recycle1198821 and parameters of feed ore size fraction
distribution 119860MF 119861MF The regression model is
[119886 120572 120583 Λ]119879
=[[
[
11990911
sdot sdot sdot 1199091119895
d
1199091198941
sdot sdot sdot 119909119894119895
]]
]
sdot [1198821 120601119861 119860MF 119861MF 119872MF]119879
(15)
The value of 119909119894119895(119894 = 1 2 3 4 119895 = 1 2 5) can be
obtained by the experimental data regression 119860MF 119861MF can
12 Journal of Applied Mathematics
be obtained from feed ore size fraction distribution and 119865(119894)is the cumulative particle percentage less than the 119894th sizefraction in feed ore and it is represented as follows
119865 (119894) = 1 minus exp (minus119860MF119889119861MF119894
) (16)
422 Spiral Classifier Model 120572119888119894is the particle percentage
of the 119894th size fraction in the first classifier overflow and 119901119894
is the particle percentage of the 119894th size fraction in ball milloverflow The spiral classifier model is as follows
120572119888119894=
119901119894times (1 minus 119864119886
119894)
sum119894=1
(119901119894times (1 minus 119864119886
119894))times 100 (17)
where 119864119886119894is the efficiency of the first spiral classifier and the
mechanism formula of 119864119886119894is shown as follows
119864119886119894= 1 minus exp [minus0693(
119889119894minus 119889min
11988950119888
minus 119889min)
119898
]
+ 119864119886min sdot [1 minus (119889119894minus 119889min119889max
)]
119896
(18)
where 119889119894is the particle with the 119894th size 119889min and 119889max
represent maximum and minimum particle sizes 11988950119888
isthe particle size fraction after correction separation 119898 isseparation accuracy and 119896 is intermix index
Through data acquisition and testing of grinding andclassification steady-state loop the regressionmodel betweenclassification parameters and condition variables size frac-tion distribution can be established The input variablesinclude the solid flow of feed ore 119872MF the classifier wateraddition 119882
2 and the parameters of ball mill overflow size
fraction distribution 119860MF 119861MF The regression model isshown as follows
[119864119886min 11988950119888
119898 119896]119879
=[[
[
11991011
sdot sdot sdot 1199101119895
d
1199101198941
sdot sdot sdot 119910119894119895
]]
]
sdot [119872MF 1198822119860MP 119861MP]
119879
(19)
where the value of 119910119894119895(119894 = 1 2 3 4 119895 = 1 2 3 4) can be
obtained by data regressionThe first-stage overflow 119875
2calculation formula is as
follows
1198752=
119872MF(119872MF +1198821 +1198822)
(20)
Similarly we can get the second spiral classifier model asfollows
1205721198881015840
119894=
120572119888119894times (1 minus 119864119886
1015840
119894)
sum119894(120572119888119894times (1 minus 1198641198861015840))
times 100 (21)
where 1205721198881015840119894is the particle percentage of the 119894th size fraction in
the second classifier overflow 120572119888119894is the particle percentage
of the 119894th size fraction in the first classifier overflow and
1198641198861015840
119894is the efficiency of the second spiral classifier The spiral
classifier model is as follows
1198641198861015840
119894= 1 minus exp[
[
minus0693(1198891015840
119894minus 1198891015840
min1198891015840
50119888minus 1198891015840
min)
1198981015840
]
]
+ 1198641198861015840
min sdot [1 minus (1198891015840
119894minus 1198891015840
min1198891015840max
)]
1198961015840
(22)
where 1198641198861015840min 1198891015840
50119888 1198981015840 and 119896
1015840 are key parameters to theefficiency of the second spiral classifier Through data acqui-sition and testing of grinding and classification steady-stateloop the regression model between classification parametersand condition variables size fraction distribution can beestablished The input variables include solid flow of feedore119872MF and parameters of the first-stage classifier overflowsize fraction distribution 119860
119888 119861119888 which are solved by similar
equation to (20) The regression model is shown as follows
[1198641198861015840
min 1198891015840
5011988811989810158401198961015840]119879
=
[[[
[
1199101015840
11sdot sdot sdot 1199101015840
1119895
d
1199101015840
1198941sdot sdot sdot 119910
1015840
119894119895
]]]
]
sdot [119872MF 119860119888119861119888]119879
(23)
where the value of1199101015840119894119895(119894 = 1 2 3 4 119895 = 1 2 3) can be obtained
by experimental data regression
43 Optimization Model of Grinding and Classification Pro-cess Two objective functions in the process are identifiedone is tomaximize output119891
1(119883) and the other is tomaximize
the small-size fractions (less than 0075mm fractions) in thesecond-stage overflow 119891
2(119883) It is also necessary to ensure
that the grinding productmeets all of other technical require-ments and the least disturbance in the following flotationcircuit As the constraints the feed load of the grinding circuit119872MF the steel ball filling rate 120601
119861 the first and the second
overflows 1198751and 119875
2 and the particle percentage of fine size
fraction in the first and the second classifier overflows 120572119888minus0075
and 1205721198881015840minus0075
should be within the user specified boundsThe operation variables are the solid flow of feed ore
119872MF water addition of the first classifier recycle 1198821 ball
filling rate 120601119861 and water addition of the second classifier119882
2
Based on all of the above grinding and classification processmultiobjective optimization model is as follows
max 119865 = max [1198911(119883) 119891
2(119883)]
1198911(119883) = 119872MF
1198912(119883) = 120572119888
1015840
minus0075
= 119891 (119872MF11988211198822 120601119861)
st 119872MFmin le 119872MF le 119872MFmax120601119861min le 120601119861 le 120601119861max1198751min le 1198751 le 1198751max1198752min le 1198752 le 1198752max120572119888minus0075
ge 120572119888min
1205721198881015840
minus0075ge 1205721198881015840
min
(24)
Journal of Applied Mathematics 13
Table 6 The optimization results calculated by the proposed algo-rithm
Number 1198911(119883) (th) 119891
2(119883) ()
1 92972 908142 91810 919003 92360 909234 90620 934605 91310 924946 90170 938007 89390 952008 89480 949329 91530 9240010 89298 9576311 89824 9454912 89800 9439513 89710 9461814 91170 9280015 91880 9184016 92190 9128117 89870 9409018 91000 9328519 91960 9152020 91124 93221
85
87
89
91
93
95
97
74000 79000 84000 94000
Optimization resultsThe raw data
f1(x)
f2(x)
()
89000
Figure 8 The comparison chart between optimization results andindustrial data
In (24) 119891(119872MF11988211198822 120601119861) implicates the model ofgrinding and classification represented by (12)ndash(23) Theproposed algorithm is applied to solve the problem and theoptimization results are shown in Table 6
With the practical process data from a grinding circuitof a mineral plant the simulation of this hybrid intelligentmethod adopted the same parameters on the variation infresh slurry feed velocity density particle size distributionand cyclone feed operating configurations
The comparison of production data and optimizationresults in Table 6 is shown in Figure 8 where ldquoXrdquo repre-sents the proposed algorithm optimization results and ldquoIrdquorepresents the original data collected from the field withoutoptimization of raw data According to the objectives thedata point closer to the upper right edge is more beneficialObviously the proposed optimization result is far betterthan the original data indicating the effectiveness of theoptimization approach
44 TOPSIS Method for Solution Selection The resolution ofamultiobjective optimization problemdoes not endwhen thePareto-optimal set is found In practical operational prob-lems a single solution must be selected TOPSIS [36] is auseful technique in dealing with multiattribute or multicrite-ria decision-making (MADMMCDM) problems in the realworld The standard TOPSIS method attempts to choosealternatives that simultaneously have the shortest distancefrom the positive-ideal solution and the farthest distancefrom the negative-ideal solution According to the TOPSISmethod the relative closeness coefficient is calculated andthe best solution in Table 6 is the solution number 10 as1198911(119883) = 89298 (119905ℎ) and 119891
2(119883) = 95763 () The
corresponding decision variables are 119872MF = 89298 (119905ℎ)120601119861= 32119882
1= 75127 (119905ℎ) and119882
2= 16296 (119905ℎ)
5 Conclusions
Promoted by the requirements of engineering optimizationin complex practical processes of grinding and classificationwe proposed a hybrid multiobjective differential evolutionalgorithm with a few beneficial features integrated Firstly anarchiving mechanism for infeasible solutions is establishedwith functional partitioned multi-population which aims todirect the population to approach or land in the feasibleregion from different directions during evolution Secondlywe propose an infeasible constraint violations function toselect infeasible population with better performance sothat they are allowed to be saved and to participate inthe subsequent evolution Thirdly a nondominating rankingstrategy is designed to improve the crowding-distance sortingand return uniform distribution of Pareto solutions Finallythe simplex method is inserted in the differential evolutionprocess to purposefully enrich the diversity without excessivecomputation cost The advantage of the proposed algorithmis the exemption from constructing penalty function andthe preservation of meaningful infeasible solutions directlySimulation results on benchmarks indicate that the proposedalgorithm can converge quickly and effectively to the truePareto frontier with better distribution
Based on the investigated information about grinding cir-cuit process we established a multiobjective optimal modelwith equations frommechanism knowledge parameters rec-ognized by data regression and constraints of technicalrequirements The nonlinear multiobjective optimizationmodel is too complicated to be solved by traditional gradient-based algorithmsThe proposed hybrid differential evolutionalgorithm is applied and tested to achieve a Pareto solution
14 Journal of Applied Mathematics
set It is proven to be valuable for operation decision makingin the industrial process and showed superiority to the oper-ation carried out in the production In fact many operatingparameters in complex processes are highly coupled andconflicting with each other The optimal operation of theentire production process is very difficult to obtain bymanualcalculation let alone the fluctuation situation of processconditions The application case indicates that the proposedmethod has good performance and is helpful to inspirefurther research on evolutionary methods for engineeringoptimization
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgments
This work is supported by the National Natural ScienceFoundation of China (61374156 61273187 and 61134006)the National Key Technology Research and DevelopmentProgram of the Ministry of Science and Technology ofChina (2012BAK09B04) the Science Fund for CreativeResearch Groups of the National Natural Science Foundationof China (61321003) the Fund for Doctor Station of theMinistry of Education (20110162130011 and 20100162120019)and theHunanProvince Science andTechnology PlanProject(2012CK4018)
References
[1] K Mitra and R Gopinath ldquoMultiobjective optimization ofan industrial grinding operation using elitist nondominatedsorting genetic algorithmrdquo Chemical Engineering Science vol59 no 2 pp 385ndash396 2004
[2] G Yu T Y Chai and X C Luo ldquoMultiobjective productionplanning optimization using hybrid evolutionary algorithmsfor mineral processingrdquo IEEE Transactions on EvolutionaryComputation vol 15 no 4 pp 487ndash514 2011
[3] T Y Ma W H Gui Y L Wang and C H Yang ldquoDynamicoptimization control for grinding and classification processrdquoControl andDecision vol 27 no 2 pp 286ndash290 2012 (Chinese)
[4] F Y Edgeworth Mathematical Physics An Essay on the Appli-cation of Mathematics to the Moral Sciences C Kegan Paul ampCompany London UK 1881
[5] V Pareto Cours drsquoEconomie Politique F Rouge LausanneSwitzerland 1896
[6] X F Chen W H Gui Y L Wang and L Cen ldquoMulti-stepoptimal control of complex process a genetic programmingstrategy and its applicationrdquo Engineering Applications of Artifi-cial Intelligence vol 17 no 5 pp 491ndash500 2004
[7] C A C Coello ldquoEvolutionary multi-objective optimizationa historical view of the fieldrdquo IEEE Computational IntelligenceMagazine vol 1 no 1 pp 28ndash36 2006
[8] K Deb A Pratap S Agarwal and T Meyarivan ldquoA fastand elitist multiobjective genetic algorithm NSGA-IIrdquo IEEETransactions on Evolutionary Computation vol 6 no 2 pp 182ndash197 2002
[9] C M Fonesca and P J Fleming ldquoGenetic algorithms formulti objective optimization formulation discussion and gen-eralizationrdquo in Proceedings of the 5th International Conferenceon Genetic Algorithms pp 415ndash423 Morgan Kaufmann SanMateo Calif USA 1993
[10] S Sengupta S Das M Nasir A V Vasilakos and W PedryczldquoEnergy-efficient differentiated coverage of dynamic objectsusing an improved evolutionary multi-objective optimizationalgorithm with fuzzy-dominancerdquo in Proceedings of IEEECongress on Evolutionary Computation pp 1ndash8 IEEE PressBrisbane Australia 2012
[11] K V Price R M Storn and J A Lampinen DifferentialEvolution a Practical Approach to Global Optimization SpringerBerlin Germany 2005
[12] W Y Gong and Z H Cai ldquoAn improved multiobjective dif-ferential evolution based on Pareto-adaptive 120576-dominance andorthogonal designrdquo European Journal of Operational Researchvol 198 no 2 pp 576ndash601 2009
[13] X H Zeng W K Wong and S Y Leung ldquoAn operatorallocation optimization model for balancing control of thehybrid assembly lines using Pareto utility discrete differentialevolution algorithmrdquo Computers and Operations Research vol39 no 5 pp 1145ndash1159 2012
[14] E Zitzler and L Thiele ldquoMultiobjective evolutionary algo-rithms a comparative case study and the strength Paretoapproachrdquo IEEETransactions onEvolutionaryComputation vol3 no 4 pp 257ndash271 1999
[15] A Al-Ani A Alsukker and R N Khushaba ldquoFeature subsetselection using differential evolution and a wheel based searchstrategyrdquo Swarm and Evolutionary Computation vol 9 pp 15ndash26 2013
[16] G Y Li and M G Liu ldquoThe summary of differential evolutionalgorithm and its improvementsrdquo in Proceedings of the 3rdInternational Conference on Advanced Computer Theory andEngineering (ICACTE rsquo10) pp V3153ndashV3156 Chengdu ChinaAugust 2010
[17] H A Abbass ldquoThe self-adaptive Pareto differential evolutionalgorithmrdquo in Proceedings of the IEEE Congress on EvolutionaryComputation pp 831ndash836 IEEE Press Honolulu Hawaii USAMay 2002
[18] N K Madavan ldquoMultiobjective optimization using a Paretodifferential evolution approachrdquo in Proceedings of the IEEECongress on Evolutionary Computation pp 1145ndash1150 IEEEPress Honolulu Hawaii USA May 2002
[19] F Xue A C Sanderson and R J Graves ldquoPareto-based multi-objective differential evolutionrdquo in Proceedings of the IEEECongress on Evolutionary Computation pp 862ndash869 IEEEPress Piscataway NJ USA December 2003
[20] T Robic and B Filipi ldquoDEMO differential evolution formultiobjective optimizationrdquo in Evolutionary Multi-CriterionOptimization pp 520ndash533 Springer Berlin Germany 2005
[21] V L Huang A K Qin P N Suganthan and M F TasgetirenldquoMulti-objective optimization based on self-adaptive differen-tial evolution algorithmrdquo in Proceedings of the IEEE Congresson Evolutionary Computation (CEC rsquo07) pp 3601ndash3608 IEEEPress Singapore September 2007
[22] V L Huang S Z Zhao R Mallipeddi and P N SuganthanldquoMulti-objective optimization using self-adaptive differentialevolution algorithmrdquo in Proceedings of the IEEE Congress onEvolutionary Computation (CEC rsquo09) pp 190ndash194 IEEE PressTrondheim Norway May 2009
Journal of Applied Mathematics 15
[23] J A Adeyemo and F A O Otieno ldquoMulti-objective differentialevolution algorithm for solving engineering problemsrdquo Journalof Applied Sciences vol 9 no 20 pp 3652ndash3661 2009
[24] Y Wang and Z Cai ldquoCombining multiobjective optimizationwith differential evolution to solve constrained optimizationproblemsrdquo IEEE Transactions on Evolutionary Computationvol 16 no 1 pp 117ndash134 2012
[25] A Homaifar C X Qi and S H Lai ldquoConstrained optimizationvia genetic algorithmsrdquo Simulation vol 62 no 4 pp 242ndash2531994
[26] J A Joines and C R Houck ldquoOn the use of non-stationarypenalty functions to solve nonlinear constrained optimizationproblems with GArsquosrdquo in Proceedings of the 1st IEEE Conferenceon Evolutionary Computation pp 579ndash584 Orlando Fla USAJune 1994
[27] F Jimenez and J L Verdegay ldquoEvolutionary techniques forconstrained optimization problemsrdquo in Proceedings of the 7thEuropean Congress on Intelligent Techniques and Soft Computing(EUFIT rsquo99) Aachen Germany 1999
[28] C Sun J Zeng and J Pan ldquoAn improved vector particleswarm optimization for constrained optimization problemsrdquoInformation Sciences vol 181 no 6 pp 1153ndash1163 2011
[29] Z Michalewicz and M Schoenauer ldquoEvolutionary algorithmsfor constrained parameter optimization problemsrdquo Evolution-ary Computation vol 4 no 1 pp 1ndash32 1996
[30] C A C Coello and A D Christiansen ldquoMoses a multiob-jective optimization tool for engineering designrdquo EngineeringOptimization vol 31 no 1ndash3 pp 337ndash368 1999
[31] Y G Woldesenbet G G Yen and B G Tessema ldquoConstrainthandling in multiobjective evolutionary optimizationrdquo IEEETransactions on Evolutionary Computation vol 13 no 3 pp514ndash525 2009
[32] J A Nelder and R Mead ldquoA simplex method for functionminimizationrdquoThe Computer Journal vol 7 no 4 pp 308ndash3131965
[33] D A van Veldhuizen and G B Lamont ldquoMultiobjective evolu-tionary algorithms analyzing the state-of-the-artrdquo EvolutionaryComputation vol 8 no 2 pp 125ndash147 2000
[34] R Farmani D A Savic andG AWalters ldquoEvolutionarymulti-objective optimization in water distribution network designrdquoEngineering Optimization vol 37 no 2 pp 167ndash183 2005
[35] K Deb and T Goel Controlled Elitist Non-Dominated SortingGenetic Algorithms for Better Convergence Springer BerlinGermany 2001
[36] Y J Lai T Y Liu and C L Hwang ldquoTopsis for MODMrdquoEuropean Journal of Operational Research vol 76 no 3 pp486ndash500 1994
Submit your manuscripts athttpwwwhindawicom
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical Problems in Engineering
Hindawi Publishing Corporationhttpwwwhindawicom
Differential EquationsInternational Journal of
Volume 2014
Applied MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Probability and StatisticsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical PhysicsAdvances in
Complex AnalysisJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
OptimizationJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
CombinatoricsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Operations ResearchAdvances in
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Function Spaces
Abstract and Applied AnalysisHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of Mathematics and Mathematical Sciences
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Algebra
Discrete Dynamics in Nature and Society
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Decision SciencesAdvances in
Discrete MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014 Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Stochastic AnalysisInternational Journal of
Journal of Applied Mathematics 3
a new individualThemutation operator of the DE algorithmis described as follows
Considering each target individual 119909119866119894 in the 119866th gener-
ation of size119873119901 a mutant individual 119909119866+1119894
is defined by
119909119866+1
119894= 119909119866
1199033+ 119865 (119909
119866
1199031minus 119909119866
1199032) (2)
where indexes 1199031 1199032 and 119903
3represent mutually different
integers that are different from 119894 and that are randomlygenerated over [1119873119901] and 119865 is the scaling factor
The simplex method proposed by Spendley Hext andHimsworth and later refined by Nelder andMead (NM) [32]is a derivative-free line-search method that is particularlydesigned for traditional unconstrained minimization scenar-ios Clearly NM method can be deemed as a direct line-search method of the steepest descent kind The ingredientsof the replacement process consist of four basic operationsreflection expansion contraction and shrinkage Throughthese operations the simplex can improve itself and approxi-mate to a local optimum point sequentially Furthermore thesimplex can vary its shape size and orientation to adapt itselfto the local contour of the objective function
21 Main Strategy of SM-DEMO The SM-DEMO algorithmis improved by the following three points compared with DE
211 Modified Selection Operation After traditional DE evo-lution the individual 119906119866+1
119894119895may violate the boundary con-
straints 119909max119894119895
and 119909min119894119895
119906119866+1119894119895
is replaced by new individual119908119866+1
119894119895being adjusted as follows
119908119866+1
119894119895
=
119909max119894119895
+ rand () lowast (119909max119894119895
minus 119906119866+1
119894119895) if (119906119866+1
119894119895gt 119909
max119894119895
)
119909min119894119895
+ rand () lowast (119909min119894119895
minus 119906119866+1
119894119895) if (119906119866+1
119894119895lt 119909
min119894119895
)
119906119866+1
119894119895 otherwise
(3)
The new population is combined with the existing parentpopulation to form a new set 119872119892 of bigger size than 119873119901A nondominated ranking of 119872119892 is performed and the 119873119901best individuals are selected This approach allows a globalnondomination checking between both the parent and thenew generation rather than only in the new generation asis done in other approaches whereas it requires additionalcomputational cost in sorting the combined
212 Nondominated Ranking Based on Euclidean DistanceThe solutions within each nondominated frontier that residein the less crowded region in the frontier are assigned a higherrank as the NSGA-II algorithm [8] developed by Deb et alindicated The crowding distance of the 119894th solution in itsfrontier (marked with solid circles) is the average side lengthof the cuboids (shown with a dashed box in Figure 1(a))The crowding-distance computation requires sorting thepopulation according to each objective function value inascending order of magnitude As shown in Figure 1 119860 and119862 are two solutions near 119861 in the same rank and 120590
0(119861) is the
crowding distance of the119861th solution traditionally calculatedas follows
1205900(119861) =
119899
sum
119895=1
10038161003816100381610038161003816119891119895(119860) minus 119891
119895(119862)
10038161003816100381610038161003816 (4)
where 119891119895(119860) 119891
119895(119862) are the objective vectors For each
objective function the boundary solutions (solutions withthe smallest and the largest function values) are assigned aninfinite distance value
A crowding-distance metric is used to estimate thedensity of solutions surrounding a particular solution in thepopulation and is obtained from the average distance of thetwo solutions on either side of the solution along each of theobjectives As shown in Figure 1119860119861119862 are the individuals ofthe generation on the same frontier and we easily know thatthe density in Figure 1(a) is better than that in Figure 1(b) Ifwe use (4) to calculate the crowding distance of 119862 we onlyknow that in Figure 1(a) it is better than in Figure 1(b) thecrowding distance of119862 in Figures 1(a) and 1(c) is equal whichis against the knowledge
To distinguish the mentioned situations we proposean improved crowding-distance metric based on Euclideandistance 119872 is the center point of the line 119860119861 119891
119895is the 119895th
objective vector and the crowding distance 120590(119861) is defined asfollows
120590 (119861) = |119860119862| minus |119861119872|
= radic
119899
sum
119895=1
[119891119895(119860) minus 119891
119895(119862)]2
minus radic
119899
sum
119895=1
119891119895(119861) minus
[119891119895(119860) + 119891
119895(119862)]
2
2
(5)
The crowded-comparison operator guides the selectionprocess at the various stages of the algorithm toward auniformly distributed Pareto-optimal frontier To carry outthe comparison we assume that every individual in thepopulation has two attributes (1) nondomination rank 119894rankand (2) crowding distance 119894distance Then a partial order isdefined as 119894 ≺ 119895 If 119894 ≺ 119895 that is between two solutionswith different nondomination ranks we prefer the solutionwith the lower (better) rank namely 119894rank ≺ 119895rank Otherwiseif both solutions belong to the same frontier that is 119894rank =
119895rank then we prefer the solution that is located in a lessercrowded region that is 119894distance ≻ 119895distance
213 Simplex Method for Local Search The simplex methodfor local search is mixed in the evolution process to enhancethe searching strategy in the optimization process The goalof integrating NM simplex method [32] and DE is to enrichthe population diversity and avoid being trapped in localminimum We apply NM simplex method operator to thepresent population when the number of iterations is greaterthan a particular value (like 119866max2) The individuals thatachieved the single extreme value in each objective functionaremarked as the initial vertex points of simplexmethodThe
4 Journal of Applied Mathematics
A
B
C
f1
f2
(a)
A
B
C
f1
f2
(b)
A
B
C
f1
f2
(c)
A
B
C
M
f1
f2
(d)
Figure 1 Crowding-distance diagram
present population is updated according to simplex methoduntil the terminal conditions are satisfied
The computation steps of the algorithm are included inSection 32
22 Evaluation Criteria Unlike the single-objective opti-mization it is more complicated for solution quality eval-uation in the case of multiobjective optimization Many ofthe suggested methods can be summarized in two typesOne is to evaluate the convergence degree by computing theproximity between the solution frontier and the actual Paretofrontier The other is to evaluate the distribution degree ofthe solutions in objective space by computing the distancesamong the individuals Here we choose both methods toevaluate the performance of the SM-DEMO algorithm
(1) Convergence Evaluation Deb et al [8] proposed thismethod in 2002 It is described as follows
120574 =1
119876(
119876
sum
119894=1
min 1003817100381710038171003817119875lowastminus 119875FT
1003817100381710038171003817) (6)
where 120574 is the extent of convergence to a known of Pareto-optimal set 119875lowast is the obtained nondomination Pareto fron-tier 119875FT is the real nondomination Pareto frontier 119875lowast minus119875FTis the Euclidean distance of 119875lowast with 119875FT and119876 is the numberof obtained solutions
(2) Distribution Degree EvaluationThenonuniformity in thedistribution is measured by SP as follows
SP = radic 1
(119876 minus 1)
119876
sum
119894=1
(119889 minus 119889119894)2
(7)
where 119889119894is the Euclidean distance among consecutive solu-
tions in the obtained nondominated set of solutions andparameter 119889 is the average distance
23 Experimental Studies Four well-known benchmark testfunctions [33] are used here to compare the performance ofSM-DEMO with NSGA-II DEMOParent These four prob-lems are called ZDT2 ZDT3 ZDT4 and ZDT6 each has twoobjective functions We describe them in Table 1
Journal of Applied Mathematics 5
Table 1 Test problems
Test problems Objective functionsmin119865(119883) = min[119891
1(119883) 119891
2(119883)]
Range of variable
ZDT21198911(119883) = 119909
1 1198912(119883) = 119892(119883)(1 minus (
1198911
119892(119883))
2
)
119892 (119883) = 1 + 9
119899
sum
119894=2
119909119894
119899 minus 1
119899 = 30
0 le 119909119894le 1
ZDT31198911(119883) = 119909
1 1198912(119883) = 119892(1 minus radic(
1198911
119892) minus (
1198911
119892) sin (10120587119891
1))
119892 (119883) = 1 + 9
119899
sum
119894=2
119909119894
119899 minus 1
119899 = 30
0 le 119909119894le 1
ZDT41198911(119883) = 119909
1 1198912(119883) = 119892(1 minus radic(
1198911
119892))
119892 (119883) = 1 + 10 (119899 minus 1) +
119899
sum
119894=2
(1199092
119894minus 10 cos (4120587119909
119894))
119899 = 10
0 le 1199091le 1
minus5 le 119909119894le 5 119894 = 2 119899
ZDT61198911(119883) = 1 minus exp (minus4119909
1) sin6 (6120587119909
1) 1198912(119883) = 119892(1 minus (
1198911
119892)
2
)
119892 (119883) = 1 + 9
119899
sum
119894=2
119909119894
(119899 minus 1)025
119899 = 10
0 le 119909119894le 1
Table 2 The performance results of the each algorithm on the test function
Test function Algorithm 120574 SP
ZDT2DEMOParent 0005120 plusmn 0000312 0000630 plusmn 0000010
NSGA-2 0007120 plusmn 0000413 0000540 plusmn 0000940
SM-DEMO 0004013 plusmn 0000230 0000423 plusmn 0000011
ZDT3DEMOParent 0009704 plusmn 0000027 0007512 plusmn 0000165
NSGA-2 0014067 plusmn 0000059 0006540 plusmn 0000124
SM-DEMO 0004704 plusmn 0000003 0004450 plusmn 0000153
ZDT4DEMOParent 2009704 plusmn 0901164 0011031 plusmn 0001104
NSGA-2 3144067 plusmn 2100740 0010122 plusmn 0000072
SM-DEMO 0874001 plusmn 0014323 0008721 plusmn 0000159
ZDT6DEMOParent 0649704 plusmn 0004912 0104520 plusmn 0015486
NSGA-2 1014067 plusmn 0010421 0007942 plusmn 0000105
SM-DEMO 0007750 plusmn 0000083 0002014 plusmn 0000117
The simulation is carried out under the environment ofIntel Pentium 4 CPU 306GHz 512MB memory WindowsXP Professional Matlab71 Initialization parameters are setas follows population size119873119901 = 100 scaling factor 119865 = 08cross rate 119862
119877= 06 maximum evolution generation 119866max =
250 and number of SM evolution iterations 119866SM = 100All of the three algorithms are real coded with equal
population size and equal maximum evolution generationEach algorithm independently runs 20 times for each testfunction Because we cannot get the real Pareto-optimal setwewill take 60 Pareto-optimal solutions obtained by the threealgorithms as a true Pareto-optimal solution set
We evaluated the algorithms based on the two perfor-mance indexes 120574 and SP Table 2 shows themean and varianceof 120574 and SP using three algorithms SM-DEMO NSGA-II and DEMOParent We can learn from Table 2 for theZDT2 function that all of the three algorithms have a good
performance while the SM-DEMO is slightly better than theother two algorithms In terms of convergences for ZDT3ZDT4 and ZDT6 which are more complex than ZDT2 SM-DEMO is significantly better thanDEMOParent andNSGA-II
Figure 2 shows a random running of SM-DEMO algo-rithm It is clear that SM-DEMO algorithm can produce agood approximation and a uniform distribution
3 Proposed Hybrid Algorithm for CMOP
The space of constrained multiobjective optimization prob-lem can be divided into the feasible solution space and theinfeasible solution space as shown in Figure 3 where 119878 isthe search space Ω is the feasible solution space and 119885 isthe infeasible solution space 119909
119894(119894 = 1 2 3 4) is the feasible
solution and 119910119894(119894 = 1 2 3 4) is the infeasible solution
6 Journal of Applied Mathematics
SM-DEMOPareto frontier
0 01 02 03 04 05 06 07 08 09 1f1
0
01
02
03
04
05
06
07
08
09
1f2
(a) ZDT2
0
02
04
06
08
1
minus08
minus06
minus04
minus02
f2
SM-DEMOPareto frontier
0 01 02 03 04 05 06 07 08 09 1f1
(b) ZDT3
SM-DEMOPareto frontier
0 01 02 03 04 05 06 07 08 09 1f1
0
01
02
03
04
05
06
07
08
09
1
f2
(c) ZDT4
SM-DEMO
0
01
02
03
04
05
06
07
08
09
1
Pareto frontier
02 03 04 05 06 07 08 09 1f1
f2
(d) ZDT6
Figure 2 SM-DEMO simulation curve
Assume that 119909lowast is the global optimal solution and 1199101is
the closest one to 119909lowast If the infeasible population 1199101is not
excluded by the evolution algorithm it is permitted to exploreboundary regions from new directions where the optimumis likely to be found
31 General Idea of the Proposed Algorithm Researchershave gradually realized the merit of infeasible solutions insearching for the global optimum in the feasible region Someinfeasible solutions with better performance are allowed to besaved Farmani et al [34] formulated a method to ensure thatinfeasible solutions with a slight violation become feasible inevolution Based on the constraints processing approach ofmultiobjective optimization problems the proposed hybridDE algorithm avoids constructing penalty function anddeleting meaningful infeasible solutions directly
Here the proposed algorithm will produce multiplegroups of functional partitions which include an evolution-ary population 119875119892 of size 119873119901 an intermediate population119872119892 to save feasible individuals an intermediate population119878119892 to save infeasible individuals a population 119875119891 to save theoptimal feasible solution found in the search process and apopulation 119875119888 to save the optimal infeasible solution Therelationship of multi-population is shown in Figure 4
With the description of (1) equality constraints are alwaystransformed into inequality constraints as |ℎ
119895(119883)| minus 120575 le 0
where 119895 = 119901 + 1 119902 and 120575 is a positive tolerance valueTo evaluate the infeasible solution the degree of constraintviolation of individual 119883 on the 119895th constraint is calculatedas follows
119881119894(119883) =
max 0 119892119894(119883) (119894 = 1 2 119901)
max 0 10038161003816100381610038161003816ℎ119895 (119883)10038161003816100381610038161003816minus 120575 (119895 = 119901 + 1 119902)
(8)
Journal of Applied Mathematics 7
Feasible solutionspace Ω
Search space S
x2
x1y1
y4
y3
y2 x3
x4
xlowast
Infeasible solutionspace Z
Figure 3 Distribution diagram of search space
Evolutionarypopulation
Pg
Intermediatepopulation
Mg
Optimal Randomly selected
Intermediatepopulation
SgInfeasible
Feasibleindividuals
individuals
Randomlyselected
Evolution
g = g + 1
g = g + 1
g = g + 1Randomly selected
feasiblesolution Pf
Optimal infeasible
solution Pc
Figure 4The relationship diagram ofmultipopulation demonstrat-ing
The final constraint violation of each individual in thepopulation can be obtained by calculating the mean of thenormalized constraint violations
In order to take advantage of the infeasible solutionswith better performance we proposed the following adaptivedifferential mutation operator to generate individual varia-tion learning from the mutation operators in DErand-to-best1bin according to rules defined by Price et al [11]Considering each individual vector 119909119866
119894 a mutant individual
119909119866+1
119894is defined by
119909119866+1
119894
= 119909119866
119894+ 1198651sdot (119909119866
1198911minus 119909119866
1199031)+ 1198652sdot (119909119866
1198912minus 119909119866
1199032) 119877
119862ge rand ( )
119909119866
119894+ 1198651sdot (119910119866
119894minus 119909119866
1199031)+ 1198652sdot (119909119866
1198911minus 119909119866
1199032) 119877
119862lt rand ( )
(9)
where 1199031and 1199032represent different integers and also different
from 119894 randomly generated over [1119873119901] 119865 is the scalingfactor 119909119866
119891119894(119894 = 1 2 119899) is randomly generated from 119875119891
119910119866
119894(119894 = 1 2 119899) is randomly generated from 119875119888 and 119877
119862is
the mutation factor as follows
119877119862= 1198771198880sdot120574119866+ const
120574119866minus1 + const (10)
where 1198771198880is the initial value of the variability factor const is
a small constant to ensure that the fractional is meaningfuland 120574119866 is defined as follows
120574119866=the number of infeasible solutions in 119875119892
119873119901 (11)
32 Framework of the Proposed Algorithm The proposedalgorithm is described as follows
Step 1 (initialization) Generate the population 119875119892 119875119891 and119875119888 of size 119873119901 119873119875
1 and 119873119875
2 Set the value of 119862
119877(crossover
probability)119866max (the number of function evaluations)119866SM(the iterative number of evolution by NM simplex method)119892 = 1 (the current generation number) and positive controlparameter for scaling the difference vectors 119865
1 1198652 Randomly
generate the parent population 119875119892 from the decision spaceSet the 119875119891 and 119875119888 and let the intermediate populations 119878119892and119872119892 be empty
Step 2 (DE reproduction) By (3) and (9) for mutationcrossover and selection an offspring 119878119892 is created Judge theconstraints of all individuals in 119875119892 In accordance with (8)we first calculate constraint violation degree 119881
119894(119883) of all of
the individuals If 119881119894(119883) = 0 the solution is feasible and
preserved to the intermediate set 119872119892 if 119881119894(119883) gt 0 the
solution is infeasible and preserved to the intermediate set 119878119892
Step 3 (simplex method) Apply NM simplex method oper-ator to the present population if 119892 ge 119866max2 Updatethe present population 119872119892 when the number of iterationsexceeds maximum iterations
Step 4 (119875119891 construction) Rank chromosomes in 119872119892 basedon (5) and generate the elitist population 119875119891 (the size is119873119901)from the ranked population119872119892
Step 5 (119875119888 construction) Add the chromosomes in 119878119892 withslight constraint violations to the 119875119888
Step 6 (mixing the population) Combine 119878119892with the existingparent population to form a new set119872119892 Remove the dupli-cate individuals in119872119892 and the existing parent population
Step 7 (evolution) Randomly choose chromosomes from119875119888 119875119892 and 119875119891 Use the adaptive differential mutation anduniform discrete crossover to obtain the offspring population119875119892 + 1
Step 8 (termination) If the stopping criterion ismet stop andoutput the best solution else go to Step 2
33 Experimental Study In this section we choose threeproblems CTP TNK and BNH as shown in Table 3 to testthe proposed method and compare the method with thecurrent CNSGA-II [35]
8 Journal of Applied Mathematics
Table 3 Test functions
Testfunction
Objective functionmin119865(119883) = min[119891
1(119883) 119891
2(119883)]
Constraints Range ofvariable
CTP
1198911(119883) = 119909
1
1198912(119883) = exp(minus
1198911(119883)
119888 (119883))
times41 +
5
sum
119894=2
[119909119894
2minus 10 cos (2120587119909
119894)]
1198921(119883) = cos (120579) (119891
2(119883) minus 119890) minus sin (120579) 119891
1(119883)
1198922(119883) = 119886
1003816100381610038161003816sin 119887120587 [sin (120579) (1198912 (119883) minus 119890)+ cos (120579) 119891
1(119883)]119888
10038161003816100381610038161003816
119889
1198921(119883) ge 119892
2(119883)
0 le 1199091le 1
minus5 le 119909119894le 5
119894 = 2 3 4 5
BNH1198911(119883) = 4119909
1
2+ 41199092
2
1198912(119883) = (119909
1minus 5)2+ (1199092minus 5)2
1198921(119883) = (119909
1minus 5)2+ 1199092
2minus 25
1198922(119883) = minus(119909
1minus 8)2+ (1199092+ 3) + 77
1198921(119883) le 0 119892
2(119883) le 0
0 le 1199091le 5
0 le 1199092le 3
TNK1198911(119883) = 119909
1
1198912(119883) = 119909
2
1198921(119883) = minus119909
1
2minus 1199092
2+ 1
+01 cos(16 arctan(11990911199092
))
1198922(119883) = (119909
1minus 05)
2+ (1199092minus 05)
1198921(119883) le 0 119892
2(119883) minus 05 le 0
0 le 119909119894le 120587
119894 = 1 2
20
15
10
5
00
Feasible solutionInfeasible solution
02 04 06 08 10f1
f2
Pareto-optimal frontier
(a) CTP1
20
15
10
05
0002 04 06 08 10
Feasible solutionInfeasible solution
Pareto-optimal frontier
f1
f2
(b) CTP2
Figure 5 CTP solution space
For CTP problem there are the six parameters 120579 119886 119887 119888119889 and 119890 that must be chosen in a way so that a portion ofthe unconstrained Pareto-optimal region is infeasible Eachconstraint is an implicit non-linear function of decisionvariables Thus it may be difficult to find a number ofsolutions on a non-linear constraint boundary We take twosets of values for six parameters in CTP problem which aredetermined as (1) CTP1 120579 = 01120587 119886 = 40 119887 = 05 119888 = 1119889 = 2 and 119890 = minus2 (2) CTP2 120579 = minus02120587 119886 = 02 119887 = 10119888 = 1 119889 = 6 and 119890 = 1 The Pareto frontiers the feasiblesolution spaces and the infeasible solution spaces are shownin Figure 5
The parameters are initialized as followsThe size of pop-ulation 119875119892 is119873119901 = 200 size of 119875119891 is119873119875
1= 150 size of 119875119888 is
1198731198752= 10 scaling factors 119865
1and 119865
2are randomly generated
within [05 1] cross rate is 119862119877= 06 maximum evolution
generation is 119866max = 300 and number of SM evolutioniterations is 119866SM = 100 All of the proposed algorithmsand CNSGA-II are real coded with equal population size andequal maximum evolution generation Each algorithm runs20 times independently for each test function
Figure 6 shows the result of a random running of theproposed algorithm and NSGA-II the smooth curve ldquomdashrdquorepresents the Pareto frontier and ldquoXrdquo stands for the solutionachieved by the proposed algorithm or NSGA-II
It is obvious that the proposed algorithm returns abetter approximation to the true Pareto-optimal frontie anda distribution of higher uniformity We also evaluated the
Journal of Applied Mathematics 9
40
35
30
25
20
15
10
050
Pareto frontierProposed algorithm
02 0301 04 05 07 0906 08 10
f2
f1
(a) CTP1 calculated by the proposed algorithm
CNSGA-IIPareto frontier
40
35
30
25
20
15
10
05
f2
f1
0 02 0301 04 05 07 0906 08 10
(b) CTP1 calculated by CNSGA-II
1211100908070605040302
f2
Pareto frontierProposed algorithm
f1
0 02 0301 04 05 07 0906 08 10
(c) CTP2 calculated by the proposed algorithm
CNSGA-IIPareto frontier
1211100908070605040302
f2
f1
0 02 0301 04 05 07 0906 08 10
(d) CTP2 calculated by CNSGA-II
50
40
30
20
10
00 50 100 150
Pareto frontierProposed algorithm
f2
f1
(e) BNH calculated by the proposed algorithm
Pareto frontier
50
40
30
20
10
00 50 100 150
f2
f1
CNSGA-II
(f) BNH calculated by CNSGA-II
Figure 6 Continued
10 Journal of Applied Mathematics
15
10
05
00 05 10 15
Pareto frontierProposed algorithm
f2
f1
(g) TNK calculated by the proposed algorithm
Pareto frontier
15
10
05
00 05 10 15
f2
f1
CNSGA-II
(h) TNK calculated by CNSGA-II
Figure 6 The comparison chart of Pareto frontier
Table 4 The comparison of performance
Test function Algorithm 120574 SP
CTP1 CNSGA-Π 0021317 plusmn 0000323 0873321 plusmn 008725
The proposed 0009836 plusmn 0000410 0567933 plusmn 001845
CTP2 CNSGA-Π 0011120 plusmn 0000753 078314 plusmn 002843
The proposed 0007013 plusmn 0000554 0296543 plusmn 000453
BNH CNSGA-Π 0014947 plusmn 0000632 0336941 plusmn 000917
The proposed 0013766 plusmn 0000043 0209321 plusmn 000561
TNK CNSGA-Π 0013235 plusmn 0000740 0464542 plusmn 000730
The proposed 0006435 plusmn 0000017 0224560 plusmn 000159
Feed water Water
Feed ore Ball mill OverflowMixing
PumpClassifierwater
additionFirst
classifierrecycle First-
overflow Spiral classifier 2Second-stage
to flotationSpiral classifier 1 Second classifier recycle(rough concentrate)
addition
pump
stageoverflow
Figure 7 Flow diagram of the grinding and classification process
algorithms based on the two criterions 120574 and SP as shownin Table 4 It can be observed from the data in Table 4that the proposed algorithm performs significantly betterthan the classical CNSGA-II algorithm in convergence anddistribution uniformityThe simulation results show that thisalgorithm can accurately converge to global Pareto solutionsand can maintain diversity of population
4 Optimization of Grinding andClassification Process
41 Bauxite Grinding and Classification Process Thegrindingand classification process is the key preparation for thebauxite mineral processing Here we consider a bauxitegrinding process in a certain mineral company with singlegrinding and two-stage classification as shown in Figure 7
The process consists in a grinding ball mill and twospiral classifiers First classifier recycle will be put back tothe ball mill for regrinding and the first-stage overflowwill be put into second spiral classifier after being mixedwith water the second classifier recycle will be prepared forBayer production as the rough concentrate and the second-stage overflow will be sent to the next flotation processThe production objectives are composed of the productionyields technically represented by the solid flow of feed oresince the process is nonstorable and the mineral processingquality represented by percentage of the small-size fractionsof mineral particles in the second-stage overflows
Journal of Applied Mathematics 11
Table 5 Notations for the model of the grinding and classificationprocess
Notation Description
119901119894
Particle percentage of the 119894th size fraction inthe ball mill overflow
119891119895
Particle percentage of the 119894th size fraction inthe Feed ore
119862 Rate of the first classifier recycle119864119886119894 The efficiency of the first spiral classifier
120591 Themean residence time1198751 The internal concentration in ball mill119872MF The solid flow of feed ore1198821 The water addition of the first classifier recycle
1198822 The classifier water addition
119887119894119895 The breakage distribution function119878119894 The breakage rate function119889119894 The particle with the 119894th size
120572119888119894
The particle percentage of the 119894th size in thefirst classifier overflow
119889119900 The unit size of the particle1198752 First-stage overflow120601119861 Ball filling rate
119860119888 119861119888
Parameters of the first-stage classifier overflowsize fraction distribution
119864119886119894 The efficiency of the first spiral classifier
11988950119888
The particle size fraction after correctionseparation
119898 The separation accuracy119896 The intermix index
1198641198861015840
min 1198891015840
50119888 1198981015840 1198961015840 The corresponding key parameters to theefficiency of the second spiral classifier
119886 120572 120583 ΛFour key parameters to control the breakagerate function
119889min 119889max Theminimum and maximum particle sizes119860MF 119861MF Parameters of feed ore size fraction distribution
119865(119894)The cumulative particle percentage less thanthe 119894th size fraction in feed ore
1205721198881015840
119894
The particle percentage of the 119894th size fractionin the second classifier overflow
1198641198861015840
119894 The efficiency of the second spiral classifier
42 Predictive Model of the Grinding and Classification Pro-cess Here we establish the mathematical predictive modelof each unit process in the bauxite grinding and classificationprocess The notations of the indexes decision variables andparameters are listed in Table 5 These notations will be usedfor the model of the grinding and classification process
421 Ball Mill Circuit Model Here 119901119894is the particle per-
centage of 119894th size fraction in the ball mill overflow 119891119895is the
particle percentage of 119894th size fraction in feed ore rate of thefirst classifier recycle 119862 is known and 119864119886
119894is the efficiency of
the first spiral classifier According to a technical report offield investigation and study we have that
119901119894(1 + 119862) =
119889119894119894119891119894+ sum119894minus1
119895=1119894gt1119889119894119895[119864119886119894119901119895(1 + 119862) + 119891
119895]
1 minus 119889119894119894119864119886119894
119889119894119895=
119890119895 119894 = 119895
119894minus1
sum
119896=119895
119888119894119896119888119895119896(119890119896minus 119890119894) 119894 gt 119895
119888119894119895=
minus
119895minus1
sum
119896=119894
119888119894119896119888119895119896 119894 lt 119895
1 119894 = 119895
1
119878119894minus 119878119895
119894minus1
sum
119896=119895
119878119896119887119894119896119888119896119895 119894 gt 119895
119890119895=
1
(1 + 05 sdot 120591119878119895) (1 + 025 sdot 120591119878
119895)2 120591 =
81198751
119872MF (12)
where 120591 is the mean residence time of minerals 1198751is the
internal concentration in ball mill and
1198751=
119872MF (1 + 119862)
119872MF (1 + 119862) +1198821 + 03 (1198821 +1198822) (13)
where119872MF (119905ℎ) is the solid flow of feed ore1198821is the water
addition of the first classifier recycle and1198822is the classifier
water addition 119887119894119895is the breakage distribution function 119878
119894
is the breakage rate function and it satisfied the followingequation
119878119894=
(119886(119889119894119889119900)120572
)
(1 + (119889119894120583)Λ
)
(14)
where 119889119894is the particle with the 119894th size 119889
119900it is a unit when
per millimeter is a unit 119889119900= 1 119889
119894= 119894 (mm) and 119886 120572 120583
and Λ are four key parameters to control the breakage ratefunction
In a concrete grinding and classification process the ballmill size is fixed and the speed of ball mill is constantThrough data acquisition and testing of grinding and clas-sification steady-state loop the regression model between 119886120572 120583 Λ and condition variables size fraction distributioncan be established The input variables are ball filling rate120601119861 solid flow of feed ore 119872MF water addition of the first
classifier recycle1198821 and parameters of feed ore size fraction
distribution 119860MF 119861MF The regression model is
[119886 120572 120583 Λ]119879
=[[
[
11990911
sdot sdot sdot 1199091119895
d
1199091198941
sdot sdot sdot 119909119894119895
]]
]
sdot [1198821 120601119861 119860MF 119861MF 119872MF]119879
(15)
The value of 119909119894119895(119894 = 1 2 3 4 119895 = 1 2 5) can be
obtained by the experimental data regression 119860MF 119861MF can
12 Journal of Applied Mathematics
be obtained from feed ore size fraction distribution and 119865(119894)is the cumulative particle percentage less than the 119894th sizefraction in feed ore and it is represented as follows
119865 (119894) = 1 minus exp (minus119860MF119889119861MF119894
) (16)
422 Spiral Classifier Model 120572119888119894is the particle percentage
of the 119894th size fraction in the first classifier overflow and 119901119894
is the particle percentage of the 119894th size fraction in ball milloverflow The spiral classifier model is as follows
120572119888119894=
119901119894times (1 minus 119864119886
119894)
sum119894=1
(119901119894times (1 minus 119864119886
119894))times 100 (17)
where 119864119886119894is the efficiency of the first spiral classifier and the
mechanism formula of 119864119886119894is shown as follows
119864119886119894= 1 minus exp [minus0693(
119889119894minus 119889min
11988950119888
minus 119889min)
119898
]
+ 119864119886min sdot [1 minus (119889119894minus 119889min119889max
)]
119896
(18)
where 119889119894is the particle with the 119894th size 119889min and 119889max
represent maximum and minimum particle sizes 11988950119888
isthe particle size fraction after correction separation 119898 isseparation accuracy and 119896 is intermix index
Through data acquisition and testing of grinding andclassification steady-state loop the regressionmodel betweenclassification parameters and condition variables size frac-tion distribution can be established The input variablesinclude the solid flow of feed ore 119872MF the classifier wateraddition 119882
2 and the parameters of ball mill overflow size
fraction distribution 119860MF 119861MF The regression model isshown as follows
[119864119886min 11988950119888
119898 119896]119879
=[[
[
11991011
sdot sdot sdot 1199101119895
d
1199101198941
sdot sdot sdot 119910119894119895
]]
]
sdot [119872MF 1198822119860MP 119861MP]
119879
(19)
where the value of 119910119894119895(119894 = 1 2 3 4 119895 = 1 2 3 4) can be
obtained by data regressionThe first-stage overflow 119875
2calculation formula is as
follows
1198752=
119872MF(119872MF +1198821 +1198822)
(20)
Similarly we can get the second spiral classifier model asfollows
1205721198881015840
119894=
120572119888119894times (1 minus 119864119886
1015840
119894)
sum119894(120572119888119894times (1 minus 1198641198861015840))
times 100 (21)
where 1205721198881015840119894is the particle percentage of the 119894th size fraction in
the second classifier overflow 120572119888119894is the particle percentage
of the 119894th size fraction in the first classifier overflow and
1198641198861015840
119894is the efficiency of the second spiral classifier The spiral
classifier model is as follows
1198641198861015840
119894= 1 minus exp[
[
minus0693(1198891015840
119894minus 1198891015840
min1198891015840
50119888minus 1198891015840
min)
1198981015840
]
]
+ 1198641198861015840
min sdot [1 minus (1198891015840
119894minus 1198891015840
min1198891015840max
)]
1198961015840
(22)
where 1198641198861015840min 1198891015840
50119888 1198981015840 and 119896
1015840 are key parameters to theefficiency of the second spiral classifier Through data acqui-sition and testing of grinding and classification steady-stateloop the regression model between classification parametersand condition variables size fraction distribution can beestablished The input variables include solid flow of feedore119872MF and parameters of the first-stage classifier overflowsize fraction distribution 119860
119888 119861119888 which are solved by similar
equation to (20) The regression model is shown as follows
[1198641198861015840
min 1198891015840
5011988811989810158401198961015840]119879
=
[[[
[
1199101015840
11sdot sdot sdot 1199101015840
1119895
d
1199101015840
1198941sdot sdot sdot 119910
1015840
119894119895
]]]
]
sdot [119872MF 119860119888119861119888]119879
(23)
where the value of1199101015840119894119895(119894 = 1 2 3 4 119895 = 1 2 3) can be obtained
by experimental data regression
43 Optimization Model of Grinding and Classification Pro-cess Two objective functions in the process are identifiedone is tomaximize output119891
1(119883) and the other is tomaximize
the small-size fractions (less than 0075mm fractions) in thesecond-stage overflow 119891
2(119883) It is also necessary to ensure
that the grinding productmeets all of other technical require-ments and the least disturbance in the following flotationcircuit As the constraints the feed load of the grinding circuit119872MF the steel ball filling rate 120601
119861 the first and the second
overflows 1198751and 119875
2 and the particle percentage of fine size
fraction in the first and the second classifier overflows 120572119888minus0075
and 1205721198881015840minus0075
should be within the user specified boundsThe operation variables are the solid flow of feed ore
119872MF water addition of the first classifier recycle 1198821 ball
filling rate 120601119861 and water addition of the second classifier119882
2
Based on all of the above grinding and classification processmultiobjective optimization model is as follows
max 119865 = max [1198911(119883) 119891
2(119883)]
1198911(119883) = 119872MF
1198912(119883) = 120572119888
1015840
minus0075
= 119891 (119872MF11988211198822 120601119861)
st 119872MFmin le 119872MF le 119872MFmax120601119861min le 120601119861 le 120601119861max1198751min le 1198751 le 1198751max1198752min le 1198752 le 1198752max120572119888minus0075
ge 120572119888min
1205721198881015840
minus0075ge 1205721198881015840
min
(24)
Journal of Applied Mathematics 13
Table 6 The optimization results calculated by the proposed algo-rithm
Number 1198911(119883) (th) 119891
2(119883) ()
1 92972 908142 91810 919003 92360 909234 90620 934605 91310 924946 90170 938007 89390 952008 89480 949329 91530 9240010 89298 9576311 89824 9454912 89800 9439513 89710 9461814 91170 9280015 91880 9184016 92190 9128117 89870 9409018 91000 9328519 91960 9152020 91124 93221
85
87
89
91
93
95
97
74000 79000 84000 94000
Optimization resultsThe raw data
f1(x)
f2(x)
()
89000
Figure 8 The comparison chart between optimization results andindustrial data
In (24) 119891(119872MF11988211198822 120601119861) implicates the model ofgrinding and classification represented by (12)ndash(23) Theproposed algorithm is applied to solve the problem and theoptimization results are shown in Table 6
With the practical process data from a grinding circuitof a mineral plant the simulation of this hybrid intelligentmethod adopted the same parameters on the variation infresh slurry feed velocity density particle size distributionand cyclone feed operating configurations
The comparison of production data and optimizationresults in Table 6 is shown in Figure 8 where ldquoXrdquo repre-sents the proposed algorithm optimization results and ldquoIrdquorepresents the original data collected from the field withoutoptimization of raw data According to the objectives thedata point closer to the upper right edge is more beneficialObviously the proposed optimization result is far betterthan the original data indicating the effectiveness of theoptimization approach
44 TOPSIS Method for Solution Selection The resolution ofamultiobjective optimization problemdoes not endwhen thePareto-optimal set is found In practical operational prob-lems a single solution must be selected TOPSIS [36] is auseful technique in dealing with multiattribute or multicrite-ria decision-making (MADMMCDM) problems in the realworld The standard TOPSIS method attempts to choosealternatives that simultaneously have the shortest distancefrom the positive-ideal solution and the farthest distancefrom the negative-ideal solution According to the TOPSISmethod the relative closeness coefficient is calculated andthe best solution in Table 6 is the solution number 10 as1198911(119883) = 89298 (119905ℎ) and 119891
2(119883) = 95763 () The
corresponding decision variables are 119872MF = 89298 (119905ℎ)120601119861= 32119882
1= 75127 (119905ℎ) and119882
2= 16296 (119905ℎ)
5 Conclusions
Promoted by the requirements of engineering optimizationin complex practical processes of grinding and classificationwe proposed a hybrid multiobjective differential evolutionalgorithm with a few beneficial features integrated Firstly anarchiving mechanism for infeasible solutions is establishedwith functional partitioned multi-population which aims todirect the population to approach or land in the feasibleregion from different directions during evolution Secondlywe propose an infeasible constraint violations function toselect infeasible population with better performance sothat they are allowed to be saved and to participate inthe subsequent evolution Thirdly a nondominating rankingstrategy is designed to improve the crowding-distance sortingand return uniform distribution of Pareto solutions Finallythe simplex method is inserted in the differential evolutionprocess to purposefully enrich the diversity without excessivecomputation cost The advantage of the proposed algorithmis the exemption from constructing penalty function andthe preservation of meaningful infeasible solutions directlySimulation results on benchmarks indicate that the proposedalgorithm can converge quickly and effectively to the truePareto frontier with better distribution
Based on the investigated information about grinding cir-cuit process we established a multiobjective optimal modelwith equations frommechanism knowledge parameters rec-ognized by data regression and constraints of technicalrequirements The nonlinear multiobjective optimizationmodel is too complicated to be solved by traditional gradient-based algorithmsThe proposed hybrid differential evolutionalgorithm is applied and tested to achieve a Pareto solution
14 Journal of Applied Mathematics
set It is proven to be valuable for operation decision makingin the industrial process and showed superiority to the oper-ation carried out in the production In fact many operatingparameters in complex processes are highly coupled andconflicting with each other The optimal operation of theentire production process is very difficult to obtain bymanualcalculation let alone the fluctuation situation of processconditions The application case indicates that the proposedmethod has good performance and is helpful to inspirefurther research on evolutionary methods for engineeringoptimization
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgments
This work is supported by the National Natural ScienceFoundation of China (61374156 61273187 and 61134006)the National Key Technology Research and DevelopmentProgram of the Ministry of Science and Technology ofChina (2012BAK09B04) the Science Fund for CreativeResearch Groups of the National Natural Science Foundationof China (61321003) the Fund for Doctor Station of theMinistry of Education (20110162130011 and 20100162120019)and theHunanProvince Science andTechnology PlanProject(2012CK4018)
References
[1] K Mitra and R Gopinath ldquoMultiobjective optimization ofan industrial grinding operation using elitist nondominatedsorting genetic algorithmrdquo Chemical Engineering Science vol59 no 2 pp 385ndash396 2004
[2] G Yu T Y Chai and X C Luo ldquoMultiobjective productionplanning optimization using hybrid evolutionary algorithmsfor mineral processingrdquo IEEE Transactions on EvolutionaryComputation vol 15 no 4 pp 487ndash514 2011
[3] T Y Ma W H Gui Y L Wang and C H Yang ldquoDynamicoptimization control for grinding and classification processrdquoControl andDecision vol 27 no 2 pp 286ndash290 2012 (Chinese)
[4] F Y Edgeworth Mathematical Physics An Essay on the Appli-cation of Mathematics to the Moral Sciences C Kegan Paul ampCompany London UK 1881
[5] V Pareto Cours drsquoEconomie Politique F Rouge LausanneSwitzerland 1896
[6] X F Chen W H Gui Y L Wang and L Cen ldquoMulti-stepoptimal control of complex process a genetic programmingstrategy and its applicationrdquo Engineering Applications of Artifi-cial Intelligence vol 17 no 5 pp 491ndash500 2004
[7] C A C Coello ldquoEvolutionary multi-objective optimizationa historical view of the fieldrdquo IEEE Computational IntelligenceMagazine vol 1 no 1 pp 28ndash36 2006
[8] K Deb A Pratap S Agarwal and T Meyarivan ldquoA fastand elitist multiobjective genetic algorithm NSGA-IIrdquo IEEETransactions on Evolutionary Computation vol 6 no 2 pp 182ndash197 2002
[9] C M Fonesca and P J Fleming ldquoGenetic algorithms formulti objective optimization formulation discussion and gen-eralizationrdquo in Proceedings of the 5th International Conferenceon Genetic Algorithms pp 415ndash423 Morgan Kaufmann SanMateo Calif USA 1993
[10] S Sengupta S Das M Nasir A V Vasilakos and W PedryczldquoEnergy-efficient differentiated coverage of dynamic objectsusing an improved evolutionary multi-objective optimizationalgorithm with fuzzy-dominancerdquo in Proceedings of IEEECongress on Evolutionary Computation pp 1ndash8 IEEE PressBrisbane Australia 2012
[11] K V Price R M Storn and J A Lampinen DifferentialEvolution a Practical Approach to Global Optimization SpringerBerlin Germany 2005
[12] W Y Gong and Z H Cai ldquoAn improved multiobjective dif-ferential evolution based on Pareto-adaptive 120576-dominance andorthogonal designrdquo European Journal of Operational Researchvol 198 no 2 pp 576ndash601 2009
[13] X H Zeng W K Wong and S Y Leung ldquoAn operatorallocation optimization model for balancing control of thehybrid assembly lines using Pareto utility discrete differentialevolution algorithmrdquo Computers and Operations Research vol39 no 5 pp 1145ndash1159 2012
[14] E Zitzler and L Thiele ldquoMultiobjective evolutionary algo-rithms a comparative case study and the strength Paretoapproachrdquo IEEETransactions onEvolutionaryComputation vol3 no 4 pp 257ndash271 1999
[15] A Al-Ani A Alsukker and R N Khushaba ldquoFeature subsetselection using differential evolution and a wheel based searchstrategyrdquo Swarm and Evolutionary Computation vol 9 pp 15ndash26 2013
[16] G Y Li and M G Liu ldquoThe summary of differential evolutionalgorithm and its improvementsrdquo in Proceedings of the 3rdInternational Conference on Advanced Computer Theory andEngineering (ICACTE rsquo10) pp V3153ndashV3156 Chengdu ChinaAugust 2010
[17] H A Abbass ldquoThe self-adaptive Pareto differential evolutionalgorithmrdquo in Proceedings of the IEEE Congress on EvolutionaryComputation pp 831ndash836 IEEE Press Honolulu Hawaii USAMay 2002
[18] N K Madavan ldquoMultiobjective optimization using a Paretodifferential evolution approachrdquo in Proceedings of the IEEECongress on Evolutionary Computation pp 1145ndash1150 IEEEPress Honolulu Hawaii USA May 2002
[19] F Xue A C Sanderson and R J Graves ldquoPareto-based multi-objective differential evolutionrdquo in Proceedings of the IEEECongress on Evolutionary Computation pp 862ndash869 IEEEPress Piscataway NJ USA December 2003
[20] T Robic and B Filipi ldquoDEMO differential evolution formultiobjective optimizationrdquo in Evolutionary Multi-CriterionOptimization pp 520ndash533 Springer Berlin Germany 2005
[21] V L Huang A K Qin P N Suganthan and M F TasgetirenldquoMulti-objective optimization based on self-adaptive differen-tial evolution algorithmrdquo in Proceedings of the IEEE Congresson Evolutionary Computation (CEC rsquo07) pp 3601ndash3608 IEEEPress Singapore September 2007
[22] V L Huang S Z Zhao R Mallipeddi and P N SuganthanldquoMulti-objective optimization using self-adaptive differentialevolution algorithmrdquo in Proceedings of the IEEE Congress onEvolutionary Computation (CEC rsquo09) pp 190ndash194 IEEE PressTrondheim Norway May 2009
Journal of Applied Mathematics 15
[23] J A Adeyemo and F A O Otieno ldquoMulti-objective differentialevolution algorithm for solving engineering problemsrdquo Journalof Applied Sciences vol 9 no 20 pp 3652ndash3661 2009
[24] Y Wang and Z Cai ldquoCombining multiobjective optimizationwith differential evolution to solve constrained optimizationproblemsrdquo IEEE Transactions on Evolutionary Computationvol 16 no 1 pp 117ndash134 2012
[25] A Homaifar C X Qi and S H Lai ldquoConstrained optimizationvia genetic algorithmsrdquo Simulation vol 62 no 4 pp 242ndash2531994
[26] J A Joines and C R Houck ldquoOn the use of non-stationarypenalty functions to solve nonlinear constrained optimizationproblems with GArsquosrdquo in Proceedings of the 1st IEEE Conferenceon Evolutionary Computation pp 579ndash584 Orlando Fla USAJune 1994
[27] F Jimenez and J L Verdegay ldquoEvolutionary techniques forconstrained optimization problemsrdquo in Proceedings of the 7thEuropean Congress on Intelligent Techniques and Soft Computing(EUFIT rsquo99) Aachen Germany 1999
[28] C Sun J Zeng and J Pan ldquoAn improved vector particleswarm optimization for constrained optimization problemsrdquoInformation Sciences vol 181 no 6 pp 1153ndash1163 2011
[29] Z Michalewicz and M Schoenauer ldquoEvolutionary algorithmsfor constrained parameter optimization problemsrdquo Evolution-ary Computation vol 4 no 1 pp 1ndash32 1996
[30] C A C Coello and A D Christiansen ldquoMoses a multiob-jective optimization tool for engineering designrdquo EngineeringOptimization vol 31 no 1ndash3 pp 337ndash368 1999
[31] Y G Woldesenbet G G Yen and B G Tessema ldquoConstrainthandling in multiobjective evolutionary optimizationrdquo IEEETransactions on Evolutionary Computation vol 13 no 3 pp514ndash525 2009
[32] J A Nelder and R Mead ldquoA simplex method for functionminimizationrdquoThe Computer Journal vol 7 no 4 pp 308ndash3131965
[33] D A van Veldhuizen and G B Lamont ldquoMultiobjective evolu-tionary algorithms analyzing the state-of-the-artrdquo EvolutionaryComputation vol 8 no 2 pp 125ndash147 2000
[34] R Farmani D A Savic andG AWalters ldquoEvolutionarymulti-objective optimization in water distribution network designrdquoEngineering Optimization vol 37 no 2 pp 167ndash183 2005
[35] K Deb and T Goel Controlled Elitist Non-Dominated SortingGenetic Algorithms for Better Convergence Springer BerlinGermany 2001
[36] Y J Lai T Y Liu and C L Hwang ldquoTopsis for MODMrdquoEuropean Journal of Operational Research vol 76 no 3 pp486ndash500 1994
Submit your manuscripts athttpwwwhindawicom
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical Problems in Engineering
Hindawi Publishing Corporationhttpwwwhindawicom
Differential EquationsInternational Journal of
Volume 2014
Applied MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Probability and StatisticsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical PhysicsAdvances in
Complex AnalysisJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
OptimizationJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
CombinatoricsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Operations ResearchAdvances in
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Function Spaces
Abstract and Applied AnalysisHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of Mathematics and Mathematical Sciences
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Algebra
Discrete Dynamics in Nature and Society
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Decision SciencesAdvances in
Discrete MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014 Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Stochastic AnalysisInternational Journal of
4 Journal of Applied Mathematics
A
B
C
f1
f2
(a)
A
B
C
f1
f2
(b)
A
B
C
f1
f2
(c)
A
B
C
M
f1
f2
(d)
Figure 1 Crowding-distance diagram
present population is updated according to simplex methoduntil the terminal conditions are satisfied
The computation steps of the algorithm are included inSection 32
22 Evaluation Criteria Unlike the single-objective opti-mization it is more complicated for solution quality eval-uation in the case of multiobjective optimization Many ofthe suggested methods can be summarized in two typesOne is to evaluate the convergence degree by computing theproximity between the solution frontier and the actual Paretofrontier The other is to evaluate the distribution degree ofthe solutions in objective space by computing the distancesamong the individuals Here we choose both methods toevaluate the performance of the SM-DEMO algorithm
(1) Convergence Evaluation Deb et al [8] proposed thismethod in 2002 It is described as follows
120574 =1
119876(
119876
sum
119894=1
min 1003817100381710038171003817119875lowastminus 119875FT
1003817100381710038171003817) (6)
where 120574 is the extent of convergence to a known of Pareto-optimal set 119875lowast is the obtained nondomination Pareto fron-tier 119875FT is the real nondomination Pareto frontier 119875lowast minus119875FTis the Euclidean distance of 119875lowast with 119875FT and119876 is the numberof obtained solutions
(2) Distribution Degree EvaluationThenonuniformity in thedistribution is measured by SP as follows
SP = radic 1
(119876 minus 1)
119876
sum
119894=1
(119889 minus 119889119894)2
(7)
where 119889119894is the Euclidean distance among consecutive solu-
tions in the obtained nondominated set of solutions andparameter 119889 is the average distance
23 Experimental Studies Four well-known benchmark testfunctions [33] are used here to compare the performance ofSM-DEMO with NSGA-II DEMOParent These four prob-lems are called ZDT2 ZDT3 ZDT4 and ZDT6 each has twoobjective functions We describe them in Table 1
Journal of Applied Mathematics 5
Table 1 Test problems
Test problems Objective functionsmin119865(119883) = min[119891
1(119883) 119891
2(119883)]
Range of variable
ZDT21198911(119883) = 119909
1 1198912(119883) = 119892(119883)(1 minus (
1198911
119892(119883))
2
)
119892 (119883) = 1 + 9
119899
sum
119894=2
119909119894
119899 minus 1
119899 = 30
0 le 119909119894le 1
ZDT31198911(119883) = 119909
1 1198912(119883) = 119892(1 minus radic(
1198911
119892) minus (
1198911
119892) sin (10120587119891
1))
119892 (119883) = 1 + 9
119899
sum
119894=2
119909119894
119899 minus 1
119899 = 30
0 le 119909119894le 1
ZDT41198911(119883) = 119909
1 1198912(119883) = 119892(1 minus radic(
1198911
119892))
119892 (119883) = 1 + 10 (119899 minus 1) +
119899
sum
119894=2
(1199092
119894minus 10 cos (4120587119909
119894))
119899 = 10
0 le 1199091le 1
minus5 le 119909119894le 5 119894 = 2 119899
ZDT61198911(119883) = 1 minus exp (minus4119909
1) sin6 (6120587119909
1) 1198912(119883) = 119892(1 minus (
1198911
119892)
2
)
119892 (119883) = 1 + 9
119899
sum
119894=2
119909119894
(119899 minus 1)025
119899 = 10
0 le 119909119894le 1
Table 2 The performance results of the each algorithm on the test function
Test function Algorithm 120574 SP
ZDT2DEMOParent 0005120 plusmn 0000312 0000630 plusmn 0000010
NSGA-2 0007120 plusmn 0000413 0000540 plusmn 0000940
SM-DEMO 0004013 plusmn 0000230 0000423 plusmn 0000011
ZDT3DEMOParent 0009704 plusmn 0000027 0007512 plusmn 0000165
NSGA-2 0014067 plusmn 0000059 0006540 plusmn 0000124
SM-DEMO 0004704 plusmn 0000003 0004450 plusmn 0000153
ZDT4DEMOParent 2009704 plusmn 0901164 0011031 plusmn 0001104
NSGA-2 3144067 plusmn 2100740 0010122 plusmn 0000072
SM-DEMO 0874001 plusmn 0014323 0008721 plusmn 0000159
ZDT6DEMOParent 0649704 plusmn 0004912 0104520 plusmn 0015486
NSGA-2 1014067 plusmn 0010421 0007942 plusmn 0000105
SM-DEMO 0007750 plusmn 0000083 0002014 plusmn 0000117
The simulation is carried out under the environment ofIntel Pentium 4 CPU 306GHz 512MB memory WindowsXP Professional Matlab71 Initialization parameters are setas follows population size119873119901 = 100 scaling factor 119865 = 08cross rate 119862
119877= 06 maximum evolution generation 119866max =
250 and number of SM evolution iterations 119866SM = 100All of the three algorithms are real coded with equal
population size and equal maximum evolution generationEach algorithm independently runs 20 times for each testfunction Because we cannot get the real Pareto-optimal setwewill take 60 Pareto-optimal solutions obtained by the threealgorithms as a true Pareto-optimal solution set
We evaluated the algorithms based on the two perfor-mance indexes 120574 and SP Table 2 shows themean and varianceof 120574 and SP using three algorithms SM-DEMO NSGA-II and DEMOParent We can learn from Table 2 for theZDT2 function that all of the three algorithms have a good
performance while the SM-DEMO is slightly better than theother two algorithms In terms of convergences for ZDT3ZDT4 and ZDT6 which are more complex than ZDT2 SM-DEMO is significantly better thanDEMOParent andNSGA-II
Figure 2 shows a random running of SM-DEMO algo-rithm It is clear that SM-DEMO algorithm can produce agood approximation and a uniform distribution
3 Proposed Hybrid Algorithm for CMOP
The space of constrained multiobjective optimization prob-lem can be divided into the feasible solution space and theinfeasible solution space as shown in Figure 3 where 119878 isthe search space Ω is the feasible solution space and 119885 isthe infeasible solution space 119909
119894(119894 = 1 2 3 4) is the feasible
solution and 119910119894(119894 = 1 2 3 4) is the infeasible solution
6 Journal of Applied Mathematics
SM-DEMOPareto frontier
0 01 02 03 04 05 06 07 08 09 1f1
0
01
02
03
04
05
06
07
08
09
1f2
(a) ZDT2
0
02
04
06
08
1
minus08
minus06
minus04
minus02
f2
SM-DEMOPareto frontier
0 01 02 03 04 05 06 07 08 09 1f1
(b) ZDT3
SM-DEMOPareto frontier
0 01 02 03 04 05 06 07 08 09 1f1
0
01
02
03
04
05
06
07
08
09
1
f2
(c) ZDT4
SM-DEMO
0
01
02
03
04
05
06
07
08
09
1
Pareto frontier
02 03 04 05 06 07 08 09 1f1
f2
(d) ZDT6
Figure 2 SM-DEMO simulation curve
Assume that 119909lowast is the global optimal solution and 1199101is
the closest one to 119909lowast If the infeasible population 1199101is not
excluded by the evolution algorithm it is permitted to exploreboundary regions from new directions where the optimumis likely to be found
31 General Idea of the Proposed Algorithm Researchershave gradually realized the merit of infeasible solutions insearching for the global optimum in the feasible region Someinfeasible solutions with better performance are allowed to besaved Farmani et al [34] formulated a method to ensure thatinfeasible solutions with a slight violation become feasible inevolution Based on the constraints processing approach ofmultiobjective optimization problems the proposed hybridDE algorithm avoids constructing penalty function anddeleting meaningful infeasible solutions directly
Here the proposed algorithm will produce multiplegroups of functional partitions which include an evolution-ary population 119875119892 of size 119873119901 an intermediate population119872119892 to save feasible individuals an intermediate population119878119892 to save infeasible individuals a population 119875119891 to save theoptimal feasible solution found in the search process and apopulation 119875119888 to save the optimal infeasible solution Therelationship of multi-population is shown in Figure 4
With the description of (1) equality constraints are alwaystransformed into inequality constraints as |ℎ
119895(119883)| minus 120575 le 0
where 119895 = 119901 + 1 119902 and 120575 is a positive tolerance valueTo evaluate the infeasible solution the degree of constraintviolation of individual 119883 on the 119895th constraint is calculatedas follows
119881119894(119883) =
max 0 119892119894(119883) (119894 = 1 2 119901)
max 0 10038161003816100381610038161003816ℎ119895 (119883)10038161003816100381610038161003816minus 120575 (119895 = 119901 + 1 119902)
(8)
Journal of Applied Mathematics 7
Feasible solutionspace Ω
Search space S
x2
x1y1
y4
y3
y2 x3
x4
xlowast
Infeasible solutionspace Z
Figure 3 Distribution diagram of search space
Evolutionarypopulation
Pg
Intermediatepopulation
Mg
Optimal Randomly selected
Intermediatepopulation
SgInfeasible
Feasibleindividuals
individuals
Randomlyselected
Evolution
g = g + 1
g = g + 1
g = g + 1Randomly selected
feasiblesolution Pf
Optimal infeasible
solution Pc
Figure 4The relationship diagram ofmultipopulation demonstrat-ing
The final constraint violation of each individual in thepopulation can be obtained by calculating the mean of thenormalized constraint violations
In order to take advantage of the infeasible solutionswith better performance we proposed the following adaptivedifferential mutation operator to generate individual varia-tion learning from the mutation operators in DErand-to-best1bin according to rules defined by Price et al [11]Considering each individual vector 119909119866
119894 a mutant individual
119909119866+1
119894is defined by
119909119866+1
119894
= 119909119866
119894+ 1198651sdot (119909119866
1198911minus 119909119866
1199031)+ 1198652sdot (119909119866
1198912minus 119909119866
1199032) 119877
119862ge rand ( )
119909119866
119894+ 1198651sdot (119910119866
119894minus 119909119866
1199031)+ 1198652sdot (119909119866
1198911minus 119909119866
1199032) 119877
119862lt rand ( )
(9)
where 1199031and 1199032represent different integers and also different
from 119894 randomly generated over [1119873119901] 119865 is the scalingfactor 119909119866
119891119894(119894 = 1 2 119899) is randomly generated from 119875119891
119910119866
119894(119894 = 1 2 119899) is randomly generated from 119875119888 and 119877
119862is
the mutation factor as follows
119877119862= 1198771198880sdot120574119866+ const
120574119866minus1 + const (10)
where 1198771198880is the initial value of the variability factor const is
a small constant to ensure that the fractional is meaningfuland 120574119866 is defined as follows
120574119866=the number of infeasible solutions in 119875119892
119873119901 (11)
32 Framework of the Proposed Algorithm The proposedalgorithm is described as follows
Step 1 (initialization) Generate the population 119875119892 119875119891 and119875119888 of size 119873119901 119873119875
1 and 119873119875
2 Set the value of 119862
119877(crossover
probability)119866max (the number of function evaluations)119866SM(the iterative number of evolution by NM simplex method)119892 = 1 (the current generation number) and positive controlparameter for scaling the difference vectors 119865
1 1198652 Randomly
generate the parent population 119875119892 from the decision spaceSet the 119875119891 and 119875119888 and let the intermediate populations 119878119892and119872119892 be empty
Step 2 (DE reproduction) By (3) and (9) for mutationcrossover and selection an offspring 119878119892 is created Judge theconstraints of all individuals in 119875119892 In accordance with (8)we first calculate constraint violation degree 119881
119894(119883) of all of
the individuals If 119881119894(119883) = 0 the solution is feasible and
preserved to the intermediate set 119872119892 if 119881119894(119883) gt 0 the
solution is infeasible and preserved to the intermediate set 119878119892
Step 3 (simplex method) Apply NM simplex method oper-ator to the present population if 119892 ge 119866max2 Updatethe present population 119872119892 when the number of iterationsexceeds maximum iterations
Step 4 (119875119891 construction) Rank chromosomes in 119872119892 basedon (5) and generate the elitist population 119875119891 (the size is119873119901)from the ranked population119872119892
Step 5 (119875119888 construction) Add the chromosomes in 119878119892 withslight constraint violations to the 119875119888
Step 6 (mixing the population) Combine 119878119892with the existingparent population to form a new set119872119892 Remove the dupli-cate individuals in119872119892 and the existing parent population
Step 7 (evolution) Randomly choose chromosomes from119875119888 119875119892 and 119875119891 Use the adaptive differential mutation anduniform discrete crossover to obtain the offspring population119875119892 + 1
Step 8 (termination) If the stopping criterion ismet stop andoutput the best solution else go to Step 2
33 Experimental Study In this section we choose threeproblems CTP TNK and BNH as shown in Table 3 to testthe proposed method and compare the method with thecurrent CNSGA-II [35]
8 Journal of Applied Mathematics
Table 3 Test functions
Testfunction
Objective functionmin119865(119883) = min[119891
1(119883) 119891
2(119883)]
Constraints Range ofvariable
CTP
1198911(119883) = 119909
1
1198912(119883) = exp(minus
1198911(119883)
119888 (119883))
times41 +
5
sum
119894=2
[119909119894
2minus 10 cos (2120587119909
119894)]
1198921(119883) = cos (120579) (119891
2(119883) minus 119890) minus sin (120579) 119891
1(119883)
1198922(119883) = 119886
1003816100381610038161003816sin 119887120587 [sin (120579) (1198912 (119883) minus 119890)+ cos (120579) 119891
1(119883)]119888
10038161003816100381610038161003816
119889
1198921(119883) ge 119892
2(119883)
0 le 1199091le 1
minus5 le 119909119894le 5
119894 = 2 3 4 5
BNH1198911(119883) = 4119909
1
2+ 41199092
2
1198912(119883) = (119909
1minus 5)2+ (1199092minus 5)2
1198921(119883) = (119909
1minus 5)2+ 1199092
2minus 25
1198922(119883) = minus(119909
1minus 8)2+ (1199092+ 3) + 77
1198921(119883) le 0 119892
2(119883) le 0
0 le 1199091le 5
0 le 1199092le 3
TNK1198911(119883) = 119909
1
1198912(119883) = 119909
2
1198921(119883) = minus119909
1
2minus 1199092
2+ 1
+01 cos(16 arctan(11990911199092
))
1198922(119883) = (119909
1minus 05)
2+ (1199092minus 05)
1198921(119883) le 0 119892
2(119883) minus 05 le 0
0 le 119909119894le 120587
119894 = 1 2
20
15
10
5
00
Feasible solutionInfeasible solution
02 04 06 08 10f1
f2
Pareto-optimal frontier
(a) CTP1
20
15
10
05
0002 04 06 08 10
Feasible solutionInfeasible solution
Pareto-optimal frontier
f1
f2
(b) CTP2
Figure 5 CTP solution space
For CTP problem there are the six parameters 120579 119886 119887 119888119889 and 119890 that must be chosen in a way so that a portion ofthe unconstrained Pareto-optimal region is infeasible Eachconstraint is an implicit non-linear function of decisionvariables Thus it may be difficult to find a number ofsolutions on a non-linear constraint boundary We take twosets of values for six parameters in CTP problem which aredetermined as (1) CTP1 120579 = 01120587 119886 = 40 119887 = 05 119888 = 1119889 = 2 and 119890 = minus2 (2) CTP2 120579 = minus02120587 119886 = 02 119887 = 10119888 = 1 119889 = 6 and 119890 = 1 The Pareto frontiers the feasiblesolution spaces and the infeasible solution spaces are shownin Figure 5
The parameters are initialized as followsThe size of pop-ulation 119875119892 is119873119901 = 200 size of 119875119891 is119873119875
1= 150 size of 119875119888 is
1198731198752= 10 scaling factors 119865
1and 119865
2are randomly generated
within [05 1] cross rate is 119862119877= 06 maximum evolution
generation is 119866max = 300 and number of SM evolutioniterations is 119866SM = 100 All of the proposed algorithmsand CNSGA-II are real coded with equal population size andequal maximum evolution generation Each algorithm runs20 times independently for each test function
Figure 6 shows the result of a random running of theproposed algorithm and NSGA-II the smooth curve ldquomdashrdquorepresents the Pareto frontier and ldquoXrdquo stands for the solutionachieved by the proposed algorithm or NSGA-II
It is obvious that the proposed algorithm returns abetter approximation to the true Pareto-optimal frontie anda distribution of higher uniformity We also evaluated the
Journal of Applied Mathematics 9
40
35
30
25
20
15
10
050
Pareto frontierProposed algorithm
02 0301 04 05 07 0906 08 10
f2
f1
(a) CTP1 calculated by the proposed algorithm
CNSGA-IIPareto frontier
40
35
30
25
20
15
10
05
f2
f1
0 02 0301 04 05 07 0906 08 10
(b) CTP1 calculated by CNSGA-II
1211100908070605040302
f2
Pareto frontierProposed algorithm
f1
0 02 0301 04 05 07 0906 08 10
(c) CTP2 calculated by the proposed algorithm
CNSGA-IIPareto frontier
1211100908070605040302
f2
f1
0 02 0301 04 05 07 0906 08 10
(d) CTP2 calculated by CNSGA-II
50
40
30
20
10
00 50 100 150
Pareto frontierProposed algorithm
f2
f1
(e) BNH calculated by the proposed algorithm
Pareto frontier
50
40
30
20
10
00 50 100 150
f2
f1
CNSGA-II
(f) BNH calculated by CNSGA-II
Figure 6 Continued
10 Journal of Applied Mathematics
15
10
05
00 05 10 15
Pareto frontierProposed algorithm
f2
f1
(g) TNK calculated by the proposed algorithm
Pareto frontier
15
10
05
00 05 10 15
f2
f1
CNSGA-II
(h) TNK calculated by CNSGA-II
Figure 6 The comparison chart of Pareto frontier
Table 4 The comparison of performance
Test function Algorithm 120574 SP
CTP1 CNSGA-Π 0021317 plusmn 0000323 0873321 plusmn 008725
The proposed 0009836 plusmn 0000410 0567933 plusmn 001845
CTP2 CNSGA-Π 0011120 plusmn 0000753 078314 plusmn 002843
The proposed 0007013 plusmn 0000554 0296543 plusmn 000453
BNH CNSGA-Π 0014947 plusmn 0000632 0336941 plusmn 000917
The proposed 0013766 plusmn 0000043 0209321 plusmn 000561
TNK CNSGA-Π 0013235 plusmn 0000740 0464542 plusmn 000730
The proposed 0006435 plusmn 0000017 0224560 plusmn 000159
Feed water Water
Feed ore Ball mill OverflowMixing
PumpClassifierwater
additionFirst
classifierrecycle First-
overflow Spiral classifier 2Second-stage
to flotationSpiral classifier 1 Second classifier recycle(rough concentrate)
addition
pump
stageoverflow
Figure 7 Flow diagram of the grinding and classification process
algorithms based on the two criterions 120574 and SP as shownin Table 4 It can be observed from the data in Table 4that the proposed algorithm performs significantly betterthan the classical CNSGA-II algorithm in convergence anddistribution uniformityThe simulation results show that thisalgorithm can accurately converge to global Pareto solutionsand can maintain diversity of population
4 Optimization of Grinding andClassification Process
41 Bauxite Grinding and Classification Process Thegrindingand classification process is the key preparation for thebauxite mineral processing Here we consider a bauxitegrinding process in a certain mineral company with singlegrinding and two-stage classification as shown in Figure 7
The process consists in a grinding ball mill and twospiral classifiers First classifier recycle will be put back tothe ball mill for regrinding and the first-stage overflowwill be put into second spiral classifier after being mixedwith water the second classifier recycle will be prepared forBayer production as the rough concentrate and the second-stage overflow will be sent to the next flotation processThe production objectives are composed of the productionyields technically represented by the solid flow of feed oresince the process is nonstorable and the mineral processingquality represented by percentage of the small-size fractionsof mineral particles in the second-stage overflows
Journal of Applied Mathematics 11
Table 5 Notations for the model of the grinding and classificationprocess
Notation Description
119901119894
Particle percentage of the 119894th size fraction inthe ball mill overflow
119891119895
Particle percentage of the 119894th size fraction inthe Feed ore
119862 Rate of the first classifier recycle119864119886119894 The efficiency of the first spiral classifier
120591 Themean residence time1198751 The internal concentration in ball mill119872MF The solid flow of feed ore1198821 The water addition of the first classifier recycle
1198822 The classifier water addition
119887119894119895 The breakage distribution function119878119894 The breakage rate function119889119894 The particle with the 119894th size
120572119888119894
The particle percentage of the 119894th size in thefirst classifier overflow
119889119900 The unit size of the particle1198752 First-stage overflow120601119861 Ball filling rate
119860119888 119861119888
Parameters of the first-stage classifier overflowsize fraction distribution
119864119886119894 The efficiency of the first spiral classifier
11988950119888
The particle size fraction after correctionseparation
119898 The separation accuracy119896 The intermix index
1198641198861015840
min 1198891015840
50119888 1198981015840 1198961015840 The corresponding key parameters to theefficiency of the second spiral classifier
119886 120572 120583 ΛFour key parameters to control the breakagerate function
119889min 119889max Theminimum and maximum particle sizes119860MF 119861MF Parameters of feed ore size fraction distribution
119865(119894)The cumulative particle percentage less thanthe 119894th size fraction in feed ore
1205721198881015840
119894
The particle percentage of the 119894th size fractionin the second classifier overflow
1198641198861015840
119894 The efficiency of the second spiral classifier
42 Predictive Model of the Grinding and Classification Pro-cess Here we establish the mathematical predictive modelof each unit process in the bauxite grinding and classificationprocess The notations of the indexes decision variables andparameters are listed in Table 5 These notations will be usedfor the model of the grinding and classification process
421 Ball Mill Circuit Model Here 119901119894is the particle per-
centage of 119894th size fraction in the ball mill overflow 119891119895is the
particle percentage of 119894th size fraction in feed ore rate of thefirst classifier recycle 119862 is known and 119864119886
119894is the efficiency of
the first spiral classifier According to a technical report offield investigation and study we have that
119901119894(1 + 119862) =
119889119894119894119891119894+ sum119894minus1
119895=1119894gt1119889119894119895[119864119886119894119901119895(1 + 119862) + 119891
119895]
1 minus 119889119894119894119864119886119894
119889119894119895=
119890119895 119894 = 119895
119894minus1
sum
119896=119895
119888119894119896119888119895119896(119890119896minus 119890119894) 119894 gt 119895
119888119894119895=
minus
119895minus1
sum
119896=119894
119888119894119896119888119895119896 119894 lt 119895
1 119894 = 119895
1
119878119894minus 119878119895
119894minus1
sum
119896=119895
119878119896119887119894119896119888119896119895 119894 gt 119895
119890119895=
1
(1 + 05 sdot 120591119878119895) (1 + 025 sdot 120591119878
119895)2 120591 =
81198751
119872MF (12)
where 120591 is the mean residence time of minerals 1198751is the
internal concentration in ball mill and
1198751=
119872MF (1 + 119862)
119872MF (1 + 119862) +1198821 + 03 (1198821 +1198822) (13)
where119872MF (119905ℎ) is the solid flow of feed ore1198821is the water
addition of the first classifier recycle and1198822is the classifier
water addition 119887119894119895is the breakage distribution function 119878
119894
is the breakage rate function and it satisfied the followingequation
119878119894=
(119886(119889119894119889119900)120572
)
(1 + (119889119894120583)Λ
)
(14)
where 119889119894is the particle with the 119894th size 119889
119900it is a unit when
per millimeter is a unit 119889119900= 1 119889
119894= 119894 (mm) and 119886 120572 120583
and Λ are four key parameters to control the breakage ratefunction
In a concrete grinding and classification process the ballmill size is fixed and the speed of ball mill is constantThrough data acquisition and testing of grinding and clas-sification steady-state loop the regression model between 119886120572 120583 Λ and condition variables size fraction distributioncan be established The input variables are ball filling rate120601119861 solid flow of feed ore 119872MF water addition of the first
classifier recycle1198821 and parameters of feed ore size fraction
distribution 119860MF 119861MF The regression model is
[119886 120572 120583 Λ]119879
=[[
[
11990911
sdot sdot sdot 1199091119895
d
1199091198941
sdot sdot sdot 119909119894119895
]]
]
sdot [1198821 120601119861 119860MF 119861MF 119872MF]119879
(15)
The value of 119909119894119895(119894 = 1 2 3 4 119895 = 1 2 5) can be
obtained by the experimental data regression 119860MF 119861MF can
12 Journal of Applied Mathematics
be obtained from feed ore size fraction distribution and 119865(119894)is the cumulative particle percentage less than the 119894th sizefraction in feed ore and it is represented as follows
119865 (119894) = 1 minus exp (minus119860MF119889119861MF119894
) (16)
422 Spiral Classifier Model 120572119888119894is the particle percentage
of the 119894th size fraction in the first classifier overflow and 119901119894
is the particle percentage of the 119894th size fraction in ball milloverflow The spiral classifier model is as follows
120572119888119894=
119901119894times (1 minus 119864119886
119894)
sum119894=1
(119901119894times (1 minus 119864119886
119894))times 100 (17)
where 119864119886119894is the efficiency of the first spiral classifier and the
mechanism formula of 119864119886119894is shown as follows
119864119886119894= 1 minus exp [minus0693(
119889119894minus 119889min
11988950119888
minus 119889min)
119898
]
+ 119864119886min sdot [1 minus (119889119894minus 119889min119889max
)]
119896
(18)
where 119889119894is the particle with the 119894th size 119889min and 119889max
represent maximum and minimum particle sizes 11988950119888
isthe particle size fraction after correction separation 119898 isseparation accuracy and 119896 is intermix index
Through data acquisition and testing of grinding andclassification steady-state loop the regressionmodel betweenclassification parameters and condition variables size frac-tion distribution can be established The input variablesinclude the solid flow of feed ore 119872MF the classifier wateraddition 119882
2 and the parameters of ball mill overflow size
fraction distribution 119860MF 119861MF The regression model isshown as follows
[119864119886min 11988950119888
119898 119896]119879
=[[
[
11991011
sdot sdot sdot 1199101119895
d
1199101198941
sdot sdot sdot 119910119894119895
]]
]
sdot [119872MF 1198822119860MP 119861MP]
119879
(19)
where the value of 119910119894119895(119894 = 1 2 3 4 119895 = 1 2 3 4) can be
obtained by data regressionThe first-stage overflow 119875
2calculation formula is as
follows
1198752=
119872MF(119872MF +1198821 +1198822)
(20)
Similarly we can get the second spiral classifier model asfollows
1205721198881015840
119894=
120572119888119894times (1 minus 119864119886
1015840
119894)
sum119894(120572119888119894times (1 minus 1198641198861015840))
times 100 (21)
where 1205721198881015840119894is the particle percentage of the 119894th size fraction in
the second classifier overflow 120572119888119894is the particle percentage
of the 119894th size fraction in the first classifier overflow and
1198641198861015840
119894is the efficiency of the second spiral classifier The spiral
classifier model is as follows
1198641198861015840
119894= 1 minus exp[
[
minus0693(1198891015840
119894minus 1198891015840
min1198891015840
50119888minus 1198891015840
min)
1198981015840
]
]
+ 1198641198861015840
min sdot [1 minus (1198891015840
119894minus 1198891015840
min1198891015840max
)]
1198961015840
(22)
where 1198641198861015840min 1198891015840
50119888 1198981015840 and 119896
1015840 are key parameters to theefficiency of the second spiral classifier Through data acqui-sition and testing of grinding and classification steady-stateloop the regression model between classification parametersand condition variables size fraction distribution can beestablished The input variables include solid flow of feedore119872MF and parameters of the first-stage classifier overflowsize fraction distribution 119860
119888 119861119888 which are solved by similar
equation to (20) The regression model is shown as follows
[1198641198861015840
min 1198891015840
5011988811989810158401198961015840]119879
=
[[[
[
1199101015840
11sdot sdot sdot 1199101015840
1119895
d
1199101015840
1198941sdot sdot sdot 119910
1015840
119894119895
]]]
]
sdot [119872MF 119860119888119861119888]119879
(23)
where the value of1199101015840119894119895(119894 = 1 2 3 4 119895 = 1 2 3) can be obtained
by experimental data regression
43 Optimization Model of Grinding and Classification Pro-cess Two objective functions in the process are identifiedone is tomaximize output119891
1(119883) and the other is tomaximize
the small-size fractions (less than 0075mm fractions) in thesecond-stage overflow 119891
2(119883) It is also necessary to ensure
that the grinding productmeets all of other technical require-ments and the least disturbance in the following flotationcircuit As the constraints the feed load of the grinding circuit119872MF the steel ball filling rate 120601
119861 the first and the second
overflows 1198751and 119875
2 and the particle percentage of fine size
fraction in the first and the second classifier overflows 120572119888minus0075
and 1205721198881015840minus0075
should be within the user specified boundsThe operation variables are the solid flow of feed ore
119872MF water addition of the first classifier recycle 1198821 ball
filling rate 120601119861 and water addition of the second classifier119882
2
Based on all of the above grinding and classification processmultiobjective optimization model is as follows
max 119865 = max [1198911(119883) 119891
2(119883)]
1198911(119883) = 119872MF
1198912(119883) = 120572119888
1015840
minus0075
= 119891 (119872MF11988211198822 120601119861)
st 119872MFmin le 119872MF le 119872MFmax120601119861min le 120601119861 le 120601119861max1198751min le 1198751 le 1198751max1198752min le 1198752 le 1198752max120572119888minus0075
ge 120572119888min
1205721198881015840
minus0075ge 1205721198881015840
min
(24)
Journal of Applied Mathematics 13
Table 6 The optimization results calculated by the proposed algo-rithm
Number 1198911(119883) (th) 119891
2(119883) ()
1 92972 908142 91810 919003 92360 909234 90620 934605 91310 924946 90170 938007 89390 952008 89480 949329 91530 9240010 89298 9576311 89824 9454912 89800 9439513 89710 9461814 91170 9280015 91880 9184016 92190 9128117 89870 9409018 91000 9328519 91960 9152020 91124 93221
85
87
89
91
93
95
97
74000 79000 84000 94000
Optimization resultsThe raw data
f1(x)
f2(x)
()
89000
Figure 8 The comparison chart between optimization results andindustrial data
In (24) 119891(119872MF11988211198822 120601119861) implicates the model ofgrinding and classification represented by (12)ndash(23) Theproposed algorithm is applied to solve the problem and theoptimization results are shown in Table 6
With the practical process data from a grinding circuitof a mineral plant the simulation of this hybrid intelligentmethod adopted the same parameters on the variation infresh slurry feed velocity density particle size distributionand cyclone feed operating configurations
The comparison of production data and optimizationresults in Table 6 is shown in Figure 8 where ldquoXrdquo repre-sents the proposed algorithm optimization results and ldquoIrdquorepresents the original data collected from the field withoutoptimization of raw data According to the objectives thedata point closer to the upper right edge is more beneficialObviously the proposed optimization result is far betterthan the original data indicating the effectiveness of theoptimization approach
44 TOPSIS Method for Solution Selection The resolution ofamultiobjective optimization problemdoes not endwhen thePareto-optimal set is found In practical operational prob-lems a single solution must be selected TOPSIS [36] is auseful technique in dealing with multiattribute or multicrite-ria decision-making (MADMMCDM) problems in the realworld The standard TOPSIS method attempts to choosealternatives that simultaneously have the shortest distancefrom the positive-ideal solution and the farthest distancefrom the negative-ideal solution According to the TOPSISmethod the relative closeness coefficient is calculated andthe best solution in Table 6 is the solution number 10 as1198911(119883) = 89298 (119905ℎ) and 119891
2(119883) = 95763 () The
corresponding decision variables are 119872MF = 89298 (119905ℎ)120601119861= 32119882
1= 75127 (119905ℎ) and119882
2= 16296 (119905ℎ)
5 Conclusions
Promoted by the requirements of engineering optimizationin complex practical processes of grinding and classificationwe proposed a hybrid multiobjective differential evolutionalgorithm with a few beneficial features integrated Firstly anarchiving mechanism for infeasible solutions is establishedwith functional partitioned multi-population which aims todirect the population to approach or land in the feasibleregion from different directions during evolution Secondlywe propose an infeasible constraint violations function toselect infeasible population with better performance sothat they are allowed to be saved and to participate inthe subsequent evolution Thirdly a nondominating rankingstrategy is designed to improve the crowding-distance sortingand return uniform distribution of Pareto solutions Finallythe simplex method is inserted in the differential evolutionprocess to purposefully enrich the diversity without excessivecomputation cost The advantage of the proposed algorithmis the exemption from constructing penalty function andthe preservation of meaningful infeasible solutions directlySimulation results on benchmarks indicate that the proposedalgorithm can converge quickly and effectively to the truePareto frontier with better distribution
Based on the investigated information about grinding cir-cuit process we established a multiobjective optimal modelwith equations frommechanism knowledge parameters rec-ognized by data regression and constraints of technicalrequirements The nonlinear multiobjective optimizationmodel is too complicated to be solved by traditional gradient-based algorithmsThe proposed hybrid differential evolutionalgorithm is applied and tested to achieve a Pareto solution
14 Journal of Applied Mathematics
set It is proven to be valuable for operation decision makingin the industrial process and showed superiority to the oper-ation carried out in the production In fact many operatingparameters in complex processes are highly coupled andconflicting with each other The optimal operation of theentire production process is very difficult to obtain bymanualcalculation let alone the fluctuation situation of processconditions The application case indicates that the proposedmethod has good performance and is helpful to inspirefurther research on evolutionary methods for engineeringoptimization
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgments
This work is supported by the National Natural ScienceFoundation of China (61374156 61273187 and 61134006)the National Key Technology Research and DevelopmentProgram of the Ministry of Science and Technology ofChina (2012BAK09B04) the Science Fund for CreativeResearch Groups of the National Natural Science Foundationof China (61321003) the Fund for Doctor Station of theMinistry of Education (20110162130011 and 20100162120019)and theHunanProvince Science andTechnology PlanProject(2012CK4018)
References
[1] K Mitra and R Gopinath ldquoMultiobjective optimization ofan industrial grinding operation using elitist nondominatedsorting genetic algorithmrdquo Chemical Engineering Science vol59 no 2 pp 385ndash396 2004
[2] G Yu T Y Chai and X C Luo ldquoMultiobjective productionplanning optimization using hybrid evolutionary algorithmsfor mineral processingrdquo IEEE Transactions on EvolutionaryComputation vol 15 no 4 pp 487ndash514 2011
[3] T Y Ma W H Gui Y L Wang and C H Yang ldquoDynamicoptimization control for grinding and classification processrdquoControl andDecision vol 27 no 2 pp 286ndash290 2012 (Chinese)
[4] F Y Edgeworth Mathematical Physics An Essay on the Appli-cation of Mathematics to the Moral Sciences C Kegan Paul ampCompany London UK 1881
[5] V Pareto Cours drsquoEconomie Politique F Rouge LausanneSwitzerland 1896
[6] X F Chen W H Gui Y L Wang and L Cen ldquoMulti-stepoptimal control of complex process a genetic programmingstrategy and its applicationrdquo Engineering Applications of Artifi-cial Intelligence vol 17 no 5 pp 491ndash500 2004
[7] C A C Coello ldquoEvolutionary multi-objective optimizationa historical view of the fieldrdquo IEEE Computational IntelligenceMagazine vol 1 no 1 pp 28ndash36 2006
[8] K Deb A Pratap S Agarwal and T Meyarivan ldquoA fastand elitist multiobjective genetic algorithm NSGA-IIrdquo IEEETransactions on Evolutionary Computation vol 6 no 2 pp 182ndash197 2002
[9] C M Fonesca and P J Fleming ldquoGenetic algorithms formulti objective optimization formulation discussion and gen-eralizationrdquo in Proceedings of the 5th International Conferenceon Genetic Algorithms pp 415ndash423 Morgan Kaufmann SanMateo Calif USA 1993
[10] S Sengupta S Das M Nasir A V Vasilakos and W PedryczldquoEnergy-efficient differentiated coverage of dynamic objectsusing an improved evolutionary multi-objective optimizationalgorithm with fuzzy-dominancerdquo in Proceedings of IEEECongress on Evolutionary Computation pp 1ndash8 IEEE PressBrisbane Australia 2012
[11] K V Price R M Storn and J A Lampinen DifferentialEvolution a Practical Approach to Global Optimization SpringerBerlin Germany 2005
[12] W Y Gong and Z H Cai ldquoAn improved multiobjective dif-ferential evolution based on Pareto-adaptive 120576-dominance andorthogonal designrdquo European Journal of Operational Researchvol 198 no 2 pp 576ndash601 2009
[13] X H Zeng W K Wong and S Y Leung ldquoAn operatorallocation optimization model for balancing control of thehybrid assembly lines using Pareto utility discrete differentialevolution algorithmrdquo Computers and Operations Research vol39 no 5 pp 1145ndash1159 2012
[14] E Zitzler and L Thiele ldquoMultiobjective evolutionary algo-rithms a comparative case study and the strength Paretoapproachrdquo IEEETransactions onEvolutionaryComputation vol3 no 4 pp 257ndash271 1999
[15] A Al-Ani A Alsukker and R N Khushaba ldquoFeature subsetselection using differential evolution and a wheel based searchstrategyrdquo Swarm and Evolutionary Computation vol 9 pp 15ndash26 2013
[16] G Y Li and M G Liu ldquoThe summary of differential evolutionalgorithm and its improvementsrdquo in Proceedings of the 3rdInternational Conference on Advanced Computer Theory andEngineering (ICACTE rsquo10) pp V3153ndashV3156 Chengdu ChinaAugust 2010
[17] H A Abbass ldquoThe self-adaptive Pareto differential evolutionalgorithmrdquo in Proceedings of the IEEE Congress on EvolutionaryComputation pp 831ndash836 IEEE Press Honolulu Hawaii USAMay 2002
[18] N K Madavan ldquoMultiobjective optimization using a Paretodifferential evolution approachrdquo in Proceedings of the IEEECongress on Evolutionary Computation pp 1145ndash1150 IEEEPress Honolulu Hawaii USA May 2002
[19] F Xue A C Sanderson and R J Graves ldquoPareto-based multi-objective differential evolutionrdquo in Proceedings of the IEEECongress on Evolutionary Computation pp 862ndash869 IEEEPress Piscataway NJ USA December 2003
[20] T Robic and B Filipi ldquoDEMO differential evolution formultiobjective optimizationrdquo in Evolutionary Multi-CriterionOptimization pp 520ndash533 Springer Berlin Germany 2005
[21] V L Huang A K Qin P N Suganthan and M F TasgetirenldquoMulti-objective optimization based on self-adaptive differen-tial evolution algorithmrdquo in Proceedings of the IEEE Congresson Evolutionary Computation (CEC rsquo07) pp 3601ndash3608 IEEEPress Singapore September 2007
[22] V L Huang S Z Zhao R Mallipeddi and P N SuganthanldquoMulti-objective optimization using self-adaptive differentialevolution algorithmrdquo in Proceedings of the IEEE Congress onEvolutionary Computation (CEC rsquo09) pp 190ndash194 IEEE PressTrondheim Norway May 2009
Journal of Applied Mathematics 15
[23] J A Adeyemo and F A O Otieno ldquoMulti-objective differentialevolution algorithm for solving engineering problemsrdquo Journalof Applied Sciences vol 9 no 20 pp 3652ndash3661 2009
[24] Y Wang and Z Cai ldquoCombining multiobjective optimizationwith differential evolution to solve constrained optimizationproblemsrdquo IEEE Transactions on Evolutionary Computationvol 16 no 1 pp 117ndash134 2012
[25] A Homaifar C X Qi and S H Lai ldquoConstrained optimizationvia genetic algorithmsrdquo Simulation vol 62 no 4 pp 242ndash2531994
[26] J A Joines and C R Houck ldquoOn the use of non-stationarypenalty functions to solve nonlinear constrained optimizationproblems with GArsquosrdquo in Proceedings of the 1st IEEE Conferenceon Evolutionary Computation pp 579ndash584 Orlando Fla USAJune 1994
[27] F Jimenez and J L Verdegay ldquoEvolutionary techniques forconstrained optimization problemsrdquo in Proceedings of the 7thEuropean Congress on Intelligent Techniques and Soft Computing(EUFIT rsquo99) Aachen Germany 1999
[28] C Sun J Zeng and J Pan ldquoAn improved vector particleswarm optimization for constrained optimization problemsrdquoInformation Sciences vol 181 no 6 pp 1153ndash1163 2011
[29] Z Michalewicz and M Schoenauer ldquoEvolutionary algorithmsfor constrained parameter optimization problemsrdquo Evolution-ary Computation vol 4 no 1 pp 1ndash32 1996
[30] C A C Coello and A D Christiansen ldquoMoses a multiob-jective optimization tool for engineering designrdquo EngineeringOptimization vol 31 no 1ndash3 pp 337ndash368 1999
[31] Y G Woldesenbet G G Yen and B G Tessema ldquoConstrainthandling in multiobjective evolutionary optimizationrdquo IEEETransactions on Evolutionary Computation vol 13 no 3 pp514ndash525 2009
[32] J A Nelder and R Mead ldquoA simplex method for functionminimizationrdquoThe Computer Journal vol 7 no 4 pp 308ndash3131965
[33] D A van Veldhuizen and G B Lamont ldquoMultiobjective evolu-tionary algorithms analyzing the state-of-the-artrdquo EvolutionaryComputation vol 8 no 2 pp 125ndash147 2000
[34] R Farmani D A Savic andG AWalters ldquoEvolutionarymulti-objective optimization in water distribution network designrdquoEngineering Optimization vol 37 no 2 pp 167ndash183 2005
[35] K Deb and T Goel Controlled Elitist Non-Dominated SortingGenetic Algorithms for Better Convergence Springer BerlinGermany 2001
[36] Y J Lai T Y Liu and C L Hwang ldquoTopsis for MODMrdquoEuropean Journal of Operational Research vol 76 no 3 pp486ndash500 1994
Submit your manuscripts athttpwwwhindawicom
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical Problems in Engineering
Hindawi Publishing Corporationhttpwwwhindawicom
Differential EquationsInternational Journal of
Volume 2014
Applied MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Probability and StatisticsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical PhysicsAdvances in
Complex AnalysisJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
OptimizationJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
CombinatoricsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Operations ResearchAdvances in
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Function Spaces
Abstract and Applied AnalysisHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of Mathematics and Mathematical Sciences
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Algebra
Discrete Dynamics in Nature and Society
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Decision SciencesAdvances in
Discrete MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014 Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Stochastic AnalysisInternational Journal of
Journal of Applied Mathematics 5
Table 1 Test problems
Test problems Objective functionsmin119865(119883) = min[119891
1(119883) 119891
2(119883)]
Range of variable
ZDT21198911(119883) = 119909
1 1198912(119883) = 119892(119883)(1 minus (
1198911
119892(119883))
2
)
119892 (119883) = 1 + 9
119899
sum
119894=2
119909119894
119899 minus 1
119899 = 30
0 le 119909119894le 1
ZDT31198911(119883) = 119909
1 1198912(119883) = 119892(1 minus radic(
1198911
119892) minus (
1198911
119892) sin (10120587119891
1))
119892 (119883) = 1 + 9
119899
sum
119894=2
119909119894
119899 minus 1
119899 = 30
0 le 119909119894le 1
ZDT41198911(119883) = 119909
1 1198912(119883) = 119892(1 minus radic(
1198911
119892))
119892 (119883) = 1 + 10 (119899 minus 1) +
119899
sum
119894=2
(1199092
119894minus 10 cos (4120587119909
119894))
119899 = 10
0 le 1199091le 1
minus5 le 119909119894le 5 119894 = 2 119899
ZDT61198911(119883) = 1 minus exp (minus4119909
1) sin6 (6120587119909
1) 1198912(119883) = 119892(1 minus (
1198911
119892)
2
)
119892 (119883) = 1 + 9
119899
sum
119894=2
119909119894
(119899 minus 1)025
119899 = 10
0 le 119909119894le 1
Table 2 The performance results of the each algorithm on the test function
Test function Algorithm 120574 SP
ZDT2DEMOParent 0005120 plusmn 0000312 0000630 plusmn 0000010
NSGA-2 0007120 plusmn 0000413 0000540 plusmn 0000940
SM-DEMO 0004013 plusmn 0000230 0000423 plusmn 0000011
ZDT3DEMOParent 0009704 plusmn 0000027 0007512 plusmn 0000165
NSGA-2 0014067 plusmn 0000059 0006540 plusmn 0000124
SM-DEMO 0004704 plusmn 0000003 0004450 plusmn 0000153
ZDT4DEMOParent 2009704 plusmn 0901164 0011031 plusmn 0001104
NSGA-2 3144067 plusmn 2100740 0010122 plusmn 0000072
SM-DEMO 0874001 plusmn 0014323 0008721 plusmn 0000159
ZDT6DEMOParent 0649704 plusmn 0004912 0104520 plusmn 0015486
NSGA-2 1014067 plusmn 0010421 0007942 plusmn 0000105
SM-DEMO 0007750 plusmn 0000083 0002014 plusmn 0000117
The simulation is carried out under the environment ofIntel Pentium 4 CPU 306GHz 512MB memory WindowsXP Professional Matlab71 Initialization parameters are setas follows population size119873119901 = 100 scaling factor 119865 = 08cross rate 119862
119877= 06 maximum evolution generation 119866max =
250 and number of SM evolution iterations 119866SM = 100All of the three algorithms are real coded with equal
population size and equal maximum evolution generationEach algorithm independently runs 20 times for each testfunction Because we cannot get the real Pareto-optimal setwewill take 60 Pareto-optimal solutions obtained by the threealgorithms as a true Pareto-optimal solution set
We evaluated the algorithms based on the two perfor-mance indexes 120574 and SP Table 2 shows themean and varianceof 120574 and SP using three algorithms SM-DEMO NSGA-II and DEMOParent We can learn from Table 2 for theZDT2 function that all of the three algorithms have a good
performance while the SM-DEMO is slightly better than theother two algorithms In terms of convergences for ZDT3ZDT4 and ZDT6 which are more complex than ZDT2 SM-DEMO is significantly better thanDEMOParent andNSGA-II
Figure 2 shows a random running of SM-DEMO algo-rithm It is clear that SM-DEMO algorithm can produce agood approximation and a uniform distribution
3 Proposed Hybrid Algorithm for CMOP
The space of constrained multiobjective optimization prob-lem can be divided into the feasible solution space and theinfeasible solution space as shown in Figure 3 where 119878 isthe search space Ω is the feasible solution space and 119885 isthe infeasible solution space 119909
119894(119894 = 1 2 3 4) is the feasible
solution and 119910119894(119894 = 1 2 3 4) is the infeasible solution
6 Journal of Applied Mathematics
SM-DEMOPareto frontier
0 01 02 03 04 05 06 07 08 09 1f1
0
01
02
03
04
05
06
07
08
09
1f2
(a) ZDT2
0
02
04
06
08
1
minus08
minus06
minus04
minus02
f2
SM-DEMOPareto frontier
0 01 02 03 04 05 06 07 08 09 1f1
(b) ZDT3
SM-DEMOPareto frontier
0 01 02 03 04 05 06 07 08 09 1f1
0
01
02
03
04
05
06
07
08
09
1
f2
(c) ZDT4
SM-DEMO
0
01
02
03
04
05
06
07
08
09
1
Pareto frontier
02 03 04 05 06 07 08 09 1f1
f2
(d) ZDT6
Figure 2 SM-DEMO simulation curve
Assume that 119909lowast is the global optimal solution and 1199101is
the closest one to 119909lowast If the infeasible population 1199101is not
excluded by the evolution algorithm it is permitted to exploreboundary regions from new directions where the optimumis likely to be found
31 General Idea of the Proposed Algorithm Researchershave gradually realized the merit of infeasible solutions insearching for the global optimum in the feasible region Someinfeasible solutions with better performance are allowed to besaved Farmani et al [34] formulated a method to ensure thatinfeasible solutions with a slight violation become feasible inevolution Based on the constraints processing approach ofmultiobjective optimization problems the proposed hybridDE algorithm avoids constructing penalty function anddeleting meaningful infeasible solutions directly
Here the proposed algorithm will produce multiplegroups of functional partitions which include an evolution-ary population 119875119892 of size 119873119901 an intermediate population119872119892 to save feasible individuals an intermediate population119878119892 to save infeasible individuals a population 119875119891 to save theoptimal feasible solution found in the search process and apopulation 119875119888 to save the optimal infeasible solution Therelationship of multi-population is shown in Figure 4
With the description of (1) equality constraints are alwaystransformed into inequality constraints as |ℎ
119895(119883)| minus 120575 le 0
where 119895 = 119901 + 1 119902 and 120575 is a positive tolerance valueTo evaluate the infeasible solution the degree of constraintviolation of individual 119883 on the 119895th constraint is calculatedas follows
119881119894(119883) =
max 0 119892119894(119883) (119894 = 1 2 119901)
max 0 10038161003816100381610038161003816ℎ119895 (119883)10038161003816100381610038161003816minus 120575 (119895 = 119901 + 1 119902)
(8)
Journal of Applied Mathematics 7
Feasible solutionspace Ω
Search space S
x2
x1y1
y4
y3
y2 x3
x4
xlowast
Infeasible solutionspace Z
Figure 3 Distribution diagram of search space
Evolutionarypopulation
Pg
Intermediatepopulation
Mg
Optimal Randomly selected
Intermediatepopulation
SgInfeasible
Feasibleindividuals
individuals
Randomlyselected
Evolution
g = g + 1
g = g + 1
g = g + 1Randomly selected
feasiblesolution Pf
Optimal infeasible
solution Pc
Figure 4The relationship diagram ofmultipopulation demonstrat-ing
The final constraint violation of each individual in thepopulation can be obtained by calculating the mean of thenormalized constraint violations
In order to take advantage of the infeasible solutionswith better performance we proposed the following adaptivedifferential mutation operator to generate individual varia-tion learning from the mutation operators in DErand-to-best1bin according to rules defined by Price et al [11]Considering each individual vector 119909119866
119894 a mutant individual
119909119866+1
119894is defined by
119909119866+1
119894
= 119909119866
119894+ 1198651sdot (119909119866
1198911minus 119909119866
1199031)+ 1198652sdot (119909119866
1198912minus 119909119866
1199032) 119877
119862ge rand ( )
119909119866
119894+ 1198651sdot (119910119866
119894minus 119909119866
1199031)+ 1198652sdot (119909119866
1198911minus 119909119866
1199032) 119877
119862lt rand ( )
(9)
where 1199031and 1199032represent different integers and also different
from 119894 randomly generated over [1119873119901] 119865 is the scalingfactor 119909119866
119891119894(119894 = 1 2 119899) is randomly generated from 119875119891
119910119866
119894(119894 = 1 2 119899) is randomly generated from 119875119888 and 119877
119862is
the mutation factor as follows
119877119862= 1198771198880sdot120574119866+ const
120574119866minus1 + const (10)
where 1198771198880is the initial value of the variability factor const is
a small constant to ensure that the fractional is meaningfuland 120574119866 is defined as follows
120574119866=the number of infeasible solutions in 119875119892
119873119901 (11)
32 Framework of the Proposed Algorithm The proposedalgorithm is described as follows
Step 1 (initialization) Generate the population 119875119892 119875119891 and119875119888 of size 119873119901 119873119875
1 and 119873119875
2 Set the value of 119862
119877(crossover
probability)119866max (the number of function evaluations)119866SM(the iterative number of evolution by NM simplex method)119892 = 1 (the current generation number) and positive controlparameter for scaling the difference vectors 119865
1 1198652 Randomly
generate the parent population 119875119892 from the decision spaceSet the 119875119891 and 119875119888 and let the intermediate populations 119878119892and119872119892 be empty
Step 2 (DE reproduction) By (3) and (9) for mutationcrossover and selection an offspring 119878119892 is created Judge theconstraints of all individuals in 119875119892 In accordance with (8)we first calculate constraint violation degree 119881
119894(119883) of all of
the individuals If 119881119894(119883) = 0 the solution is feasible and
preserved to the intermediate set 119872119892 if 119881119894(119883) gt 0 the
solution is infeasible and preserved to the intermediate set 119878119892
Step 3 (simplex method) Apply NM simplex method oper-ator to the present population if 119892 ge 119866max2 Updatethe present population 119872119892 when the number of iterationsexceeds maximum iterations
Step 4 (119875119891 construction) Rank chromosomes in 119872119892 basedon (5) and generate the elitist population 119875119891 (the size is119873119901)from the ranked population119872119892
Step 5 (119875119888 construction) Add the chromosomes in 119878119892 withslight constraint violations to the 119875119888
Step 6 (mixing the population) Combine 119878119892with the existingparent population to form a new set119872119892 Remove the dupli-cate individuals in119872119892 and the existing parent population
Step 7 (evolution) Randomly choose chromosomes from119875119888 119875119892 and 119875119891 Use the adaptive differential mutation anduniform discrete crossover to obtain the offspring population119875119892 + 1
Step 8 (termination) If the stopping criterion ismet stop andoutput the best solution else go to Step 2
33 Experimental Study In this section we choose threeproblems CTP TNK and BNH as shown in Table 3 to testthe proposed method and compare the method with thecurrent CNSGA-II [35]
8 Journal of Applied Mathematics
Table 3 Test functions
Testfunction
Objective functionmin119865(119883) = min[119891
1(119883) 119891
2(119883)]
Constraints Range ofvariable
CTP
1198911(119883) = 119909
1
1198912(119883) = exp(minus
1198911(119883)
119888 (119883))
times41 +
5
sum
119894=2
[119909119894
2minus 10 cos (2120587119909
119894)]
1198921(119883) = cos (120579) (119891
2(119883) minus 119890) minus sin (120579) 119891
1(119883)
1198922(119883) = 119886
1003816100381610038161003816sin 119887120587 [sin (120579) (1198912 (119883) minus 119890)+ cos (120579) 119891
1(119883)]119888
10038161003816100381610038161003816
119889
1198921(119883) ge 119892
2(119883)
0 le 1199091le 1
minus5 le 119909119894le 5
119894 = 2 3 4 5
BNH1198911(119883) = 4119909
1
2+ 41199092
2
1198912(119883) = (119909
1minus 5)2+ (1199092minus 5)2
1198921(119883) = (119909
1minus 5)2+ 1199092
2minus 25
1198922(119883) = minus(119909
1minus 8)2+ (1199092+ 3) + 77
1198921(119883) le 0 119892
2(119883) le 0
0 le 1199091le 5
0 le 1199092le 3
TNK1198911(119883) = 119909
1
1198912(119883) = 119909
2
1198921(119883) = minus119909
1
2minus 1199092
2+ 1
+01 cos(16 arctan(11990911199092
))
1198922(119883) = (119909
1minus 05)
2+ (1199092minus 05)
1198921(119883) le 0 119892
2(119883) minus 05 le 0
0 le 119909119894le 120587
119894 = 1 2
20
15
10
5
00
Feasible solutionInfeasible solution
02 04 06 08 10f1
f2
Pareto-optimal frontier
(a) CTP1
20
15
10
05
0002 04 06 08 10
Feasible solutionInfeasible solution
Pareto-optimal frontier
f1
f2
(b) CTP2
Figure 5 CTP solution space
For CTP problem there are the six parameters 120579 119886 119887 119888119889 and 119890 that must be chosen in a way so that a portion ofthe unconstrained Pareto-optimal region is infeasible Eachconstraint is an implicit non-linear function of decisionvariables Thus it may be difficult to find a number ofsolutions on a non-linear constraint boundary We take twosets of values for six parameters in CTP problem which aredetermined as (1) CTP1 120579 = 01120587 119886 = 40 119887 = 05 119888 = 1119889 = 2 and 119890 = minus2 (2) CTP2 120579 = minus02120587 119886 = 02 119887 = 10119888 = 1 119889 = 6 and 119890 = 1 The Pareto frontiers the feasiblesolution spaces and the infeasible solution spaces are shownin Figure 5
The parameters are initialized as followsThe size of pop-ulation 119875119892 is119873119901 = 200 size of 119875119891 is119873119875
1= 150 size of 119875119888 is
1198731198752= 10 scaling factors 119865
1and 119865
2are randomly generated
within [05 1] cross rate is 119862119877= 06 maximum evolution
generation is 119866max = 300 and number of SM evolutioniterations is 119866SM = 100 All of the proposed algorithmsand CNSGA-II are real coded with equal population size andequal maximum evolution generation Each algorithm runs20 times independently for each test function
Figure 6 shows the result of a random running of theproposed algorithm and NSGA-II the smooth curve ldquomdashrdquorepresents the Pareto frontier and ldquoXrdquo stands for the solutionachieved by the proposed algorithm or NSGA-II
It is obvious that the proposed algorithm returns abetter approximation to the true Pareto-optimal frontie anda distribution of higher uniformity We also evaluated the
Journal of Applied Mathematics 9
40
35
30
25
20
15
10
050
Pareto frontierProposed algorithm
02 0301 04 05 07 0906 08 10
f2
f1
(a) CTP1 calculated by the proposed algorithm
CNSGA-IIPareto frontier
40
35
30
25
20
15
10
05
f2
f1
0 02 0301 04 05 07 0906 08 10
(b) CTP1 calculated by CNSGA-II
1211100908070605040302
f2
Pareto frontierProposed algorithm
f1
0 02 0301 04 05 07 0906 08 10
(c) CTP2 calculated by the proposed algorithm
CNSGA-IIPareto frontier
1211100908070605040302
f2
f1
0 02 0301 04 05 07 0906 08 10
(d) CTP2 calculated by CNSGA-II
50
40
30
20
10
00 50 100 150
Pareto frontierProposed algorithm
f2
f1
(e) BNH calculated by the proposed algorithm
Pareto frontier
50
40
30
20
10
00 50 100 150
f2
f1
CNSGA-II
(f) BNH calculated by CNSGA-II
Figure 6 Continued
10 Journal of Applied Mathematics
15
10
05
00 05 10 15
Pareto frontierProposed algorithm
f2
f1
(g) TNK calculated by the proposed algorithm
Pareto frontier
15
10
05
00 05 10 15
f2
f1
CNSGA-II
(h) TNK calculated by CNSGA-II
Figure 6 The comparison chart of Pareto frontier
Table 4 The comparison of performance
Test function Algorithm 120574 SP
CTP1 CNSGA-Π 0021317 plusmn 0000323 0873321 plusmn 008725
The proposed 0009836 plusmn 0000410 0567933 plusmn 001845
CTP2 CNSGA-Π 0011120 plusmn 0000753 078314 plusmn 002843
The proposed 0007013 plusmn 0000554 0296543 plusmn 000453
BNH CNSGA-Π 0014947 plusmn 0000632 0336941 plusmn 000917
The proposed 0013766 plusmn 0000043 0209321 plusmn 000561
TNK CNSGA-Π 0013235 plusmn 0000740 0464542 plusmn 000730
The proposed 0006435 plusmn 0000017 0224560 plusmn 000159
Feed water Water
Feed ore Ball mill OverflowMixing
PumpClassifierwater
additionFirst
classifierrecycle First-
overflow Spiral classifier 2Second-stage
to flotationSpiral classifier 1 Second classifier recycle(rough concentrate)
addition
pump
stageoverflow
Figure 7 Flow diagram of the grinding and classification process
algorithms based on the two criterions 120574 and SP as shownin Table 4 It can be observed from the data in Table 4that the proposed algorithm performs significantly betterthan the classical CNSGA-II algorithm in convergence anddistribution uniformityThe simulation results show that thisalgorithm can accurately converge to global Pareto solutionsand can maintain diversity of population
4 Optimization of Grinding andClassification Process
41 Bauxite Grinding and Classification Process Thegrindingand classification process is the key preparation for thebauxite mineral processing Here we consider a bauxitegrinding process in a certain mineral company with singlegrinding and two-stage classification as shown in Figure 7
The process consists in a grinding ball mill and twospiral classifiers First classifier recycle will be put back tothe ball mill for regrinding and the first-stage overflowwill be put into second spiral classifier after being mixedwith water the second classifier recycle will be prepared forBayer production as the rough concentrate and the second-stage overflow will be sent to the next flotation processThe production objectives are composed of the productionyields technically represented by the solid flow of feed oresince the process is nonstorable and the mineral processingquality represented by percentage of the small-size fractionsof mineral particles in the second-stage overflows
Journal of Applied Mathematics 11
Table 5 Notations for the model of the grinding and classificationprocess
Notation Description
119901119894
Particle percentage of the 119894th size fraction inthe ball mill overflow
119891119895
Particle percentage of the 119894th size fraction inthe Feed ore
119862 Rate of the first classifier recycle119864119886119894 The efficiency of the first spiral classifier
120591 Themean residence time1198751 The internal concentration in ball mill119872MF The solid flow of feed ore1198821 The water addition of the first classifier recycle
1198822 The classifier water addition
119887119894119895 The breakage distribution function119878119894 The breakage rate function119889119894 The particle with the 119894th size
120572119888119894
The particle percentage of the 119894th size in thefirst classifier overflow
119889119900 The unit size of the particle1198752 First-stage overflow120601119861 Ball filling rate
119860119888 119861119888
Parameters of the first-stage classifier overflowsize fraction distribution
119864119886119894 The efficiency of the first spiral classifier
11988950119888
The particle size fraction after correctionseparation
119898 The separation accuracy119896 The intermix index
1198641198861015840
min 1198891015840
50119888 1198981015840 1198961015840 The corresponding key parameters to theefficiency of the second spiral classifier
119886 120572 120583 ΛFour key parameters to control the breakagerate function
119889min 119889max Theminimum and maximum particle sizes119860MF 119861MF Parameters of feed ore size fraction distribution
119865(119894)The cumulative particle percentage less thanthe 119894th size fraction in feed ore
1205721198881015840
119894
The particle percentage of the 119894th size fractionin the second classifier overflow
1198641198861015840
119894 The efficiency of the second spiral classifier
42 Predictive Model of the Grinding and Classification Pro-cess Here we establish the mathematical predictive modelof each unit process in the bauxite grinding and classificationprocess The notations of the indexes decision variables andparameters are listed in Table 5 These notations will be usedfor the model of the grinding and classification process
421 Ball Mill Circuit Model Here 119901119894is the particle per-
centage of 119894th size fraction in the ball mill overflow 119891119895is the
particle percentage of 119894th size fraction in feed ore rate of thefirst classifier recycle 119862 is known and 119864119886
119894is the efficiency of
the first spiral classifier According to a technical report offield investigation and study we have that
119901119894(1 + 119862) =
119889119894119894119891119894+ sum119894minus1
119895=1119894gt1119889119894119895[119864119886119894119901119895(1 + 119862) + 119891
119895]
1 minus 119889119894119894119864119886119894
119889119894119895=
119890119895 119894 = 119895
119894minus1
sum
119896=119895
119888119894119896119888119895119896(119890119896minus 119890119894) 119894 gt 119895
119888119894119895=
minus
119895minus1
sum
119896=119894
119888119894119896119888119895119896 119894 lt 119895
1 119894 = 119895
1
119878119894minus 119878119895
119894minus1
sum
119896=119895
119878119896119887119894119896119888119896119895 119894 gt 119895
119890119895=
1
(1 + 05 sdot 120591119878119895) (1 + 025 sdot 120591119878
119895)2 120591 =
81198751
119872MF (12)
where 120591 is the mean residence time of minerals 1198751is the
internal concentration in ball mill and
1198751=
119872MF (1 + 119862)
119872MF (1 + 119862) +1198821 + 03 (1198821 +1198822) (13)
where119872MF (119905ℎ) is the solid flow of feed ore1198821is the water
addition of the first classifier recycle and1198822is the classifier
water addition 119887119894119895is the breakage distribution function 119878
119894
is the breakage rate function and it satisfied the followingequation
119878119894=
(119886(119889119894119889119900)120572
)
(1 + (119889119894120583)Λ
)
(14)
where 119889119894is the particle with the 119894th size 119889
119900it is a unit when
per millimeter is a unit 119889119900= 1 119889
119894= 119894 (mm) and 119886 120572 120583
and Λ are four key parameters to control the breakage ratefunction
In a concrete grinding and classification process the ballmill size is fixed and the speed of ball mill is constantThrough data acquisition and testing of grinding and clas-sification steady-state loop the regression model between 119886120572 120583 Λ and condition variables size fraction distributioncan be established The input variables are ball filling rate120601119861 solid flow of feed ore 119872MF water addition of the first
classifier recycle1198821 and parameters of feed ore size fraction
distribution 119860MF 119861MF The regression model is
[119886 120572 120583 Λ]119879
=[[
[
11990911
sdot sdot sdot 1199091119895
d
1199091198941
sdot sdot sdot 119909119894119895
]]
]
sdot [1198821 120601119861 119860MF 119861MF 119872MF]119879
(15)
The value of 119909119894119895(119894 = 1 2 3 4 119895 = 1 2 5) can be
obtained by the experimental data regression 119860MF 119861MF can
12 Journal of Applied Mathematics
be obtained from feed ore size fraction distribution and 119865(119894)is the cumulative particle percentage less than the 119894th sizefraction in feed ore and it is represented as follows
119865 (119894) = 1 minus exp (minus119860MF119889119861MF119894
) (16)
422 Spiral Classifier Model 120572119888119894is the particle percentage
of the 119894th size fraction in the first classifier overflow and 119901119894
is the particle percentage of the 119894th size fraction in ball milloverflow The spiral classifier model is as follows
120572119888119894=
119901119894times (1 minus 119864119886
119894)
sum119894=1
(119901119894times (1 minus 119864119886
119894))times 100 (17)
where 119864119886119894is the efficiency of the first spiral classifier and the
mechanism formula of 119864119886119894is shown as follows
119864119886119894= 1 minus exp [minus0693(
119889119894minus 119889min
11988950119888
minus 119889min)
119898
]
+ 119864119886min sdot [1 minus (119889119894minus 119889min119889max
)]
119896
(18)
where 119889119894is the particle with the 119894th size 119889min and 119889max
represent maximum and minimum particle sizes 11988950119888
isthe particle size fraction after correction separation 119898 isseparation accuracy and 119896 is intermix index
Through data acquisition and testing of grinding andclassification steady-state loop the regressionmodel betweenclassification parameters and condition variables size frac-tion distribution can be established The input variablesinclude the solid flow of feed ore 119872MF the classifier wateraddition 119882
2 and the parameters of ball mill overflow size
fraction distribution 119860MF 119861MF The regression model isshown as follows
[119864119886min 11988950119888
119898 119896]119879
=[[
[
11991011
sdot sdot sdot 1199101119895
d
1199101198941
sdot sdot sdot 119910119894119895
]]
]
sdot [119872MF 1198822119860MP 119861MP]
119879
(19)
where the value of 119910119894119895(119894 = 1 2 3 4 119895 = 1 2 3 4) can be
obtained by data regressionThe first-stage overflow 119875
2calculation formula is as
follows
1198752=
119872MF(119872MF +1198821 +1198822)
(20)
Similarly we can get the second spiral classifier model asfollows
1205721198881015840
119894=
120572119888119894times (1 minus 119864119886
1015840
119894)
sum119894(120572119888119894times (1 minus 1198641198861015840))
times 100 (21)
where 1205721198881015840119894is the particle percentage of the 119894th size fraction in
the second classifier overflow 120572119888119894is the particle percentage
of the 119894th size fraction in the first classifier overflow and
1198641198861015840
119894is the efficiency of the second spiral classifier The spiral
classifier model is as follows
1198641198861015840
119894= 1 minus exp[
[
minus0693(1198891015840
119894minus 1198891015840
min1198891015840
50119888minus 1198891015840
min)
1198981015840
]
]
+ 1198641198861015840
min sdot [1 minus (1198891015840
119894minus 1198891015840
min1198891015840max
)]
1198961015840
(22)
where 1198641198861015840min 1198891015840
50119888 1198981015840 and 119896
1015840 are key parameters to theefficiency of the second spiral classifier Through data acqui-sition and testing of grinding and classification steady-stateloop the regression model between classification parametersand condition variables size fraction distribution can beestablished The input variables include solid flow of feedore119872MF and parameters of the first-stage classifier overflowsize fraction distribution 119860
119888 119861119888 which are solved by similar
equation to (20) The regression model is shown as follows
[1198641198861015840
min 1198891015840
5011988811989810158401198961015840]119879
=
[[[
[
1199101015840
11sdot sdot sdot 1199101015840
1119895
d
1199101015840
1198941sdot sdot sdot 119910
1015840
119894119895
]]]
]
sdot [119872MF 119860119888119861119888]119879
(23)
where the value of1199101015840119894119895(119894 = 1 2 3 4 119895 = 1 2 3) can be obtained
by experimental data regression
43 Optimization Model of Grinding and Classification Pro-cess Two objective functions in the process are identifiedone is tomaximize output119891
1(119883) and the other is tomaximize
the small-size fractions (less than 0075mm fractions) in thesecond-stage overflow 119891
2(119883) It is also necessary to ensure
that the grinding productmeets all of other technical require-ments and the least disturbance in the following flotationcircuit As the constraints the feed load of the grinding circuit119872MF the steel ball filling rate 120601
119861 the first and the second
overflows 1198751and 119875
2 and the particle percentage of fine size
fraction in the first and the second classifier overflows 120572119888minus0075
and 1205721198881015840minus0075
should be within the user specified boundsThe operation variables are the solid flow of feed ore
119872MF water addition of the first classifier recycle 1198821 ball
filling rate 120601119861 and water addition of the second classifier119882
2
Based on all of the above grinding and classification processmultiobjective optimization model is as follows
max 119865 = max [1198911(119883) 119891
2(119883)]
1198911(119883) = 119872MF
1198912(119883) = 120572119888
1015840
minus0075
= 119891 (119872MF11988211198822 120601119861)
st 119872MFmin le 119872MF le 119872MFmax120601119861min le 120601119861 le 120601119861max1198751min le 1198751 le 1198751max1198752min le 1198752 le 1198752max120572119888minus0075
ge 120572119888min
1205721198881015840
minus0075ge 1205721198881015840
min
(24)
Journal of Applied Mathematics 13
Table 6 The optimization results calculated by the proposed algo-rithm
Number 1198911(119883) (th) 119891
2(119883) ()
1 92972 908142 91810 919003 92360 909234 90620 934605 91310 924946 90170 938007 89390 952008 89480 949329 91530 9240010 89298 9576311 89824 9454912 89800 9439513 89710 9461814 91170 9280015 91880 9184016 92190 9128117 89870 9409018 91000 9328519 91960 9152020 91124 93221
85
87
89
91
93
95
97
74000 79000 84000 94000
Optimization resultsThe raw data
f1(x)
f2(x)
()
89000
Figure 8 The comparison chart between optimization results andindustrial data
In (24) 119891(119872MF11988211198822 120601119861) implicates the model ofgrinding and classification represented by (12)ndash(23) Theproposed algorithm is applied to solve the problem and theoptimization results are shown in Table 6
With the practical process data from a grinding circuitof a mineral plant the simulation of this hybrid intelligentmethod adopted the same parameters on the variation infresh slurry feed velocity density particle size distributionand cyclone feed operating configurations
The comparison of production data and optimizationresults in Table 6 is shown in Figure 8 where ldquoXrdquo repre-sents the proposed algorithm optimization results and ldquoIrdquorepresents the original data collected from the field withoutoptimization of raw data According to the objectives thedata point closer to the upper right edge is more beneficialObviously the proposed optimization result is far betterthan the original data indicating the effectiveness of theoptimization approach
44 TOPSIS Method for Solution Selection The resolution ofamultiobjective optimization problemdoes not endwhen thePareto-optimal set is found In practical operational prob-lems a single solution must be selected TOPSIS [36] is auseful technique in dealing with multiattribute or multicrite-ria decision-making (MADMMCDM) problems in the realworld The standard TOPSIS method attempts to choosealternatives that simultaneously have the shortest distancefrom the positive-ideal solution and the farthest distancefrom the negative-ideal solution According to the TOPSISmethod the relative closeness coefficient is calculated andthe best solution in Table 6 is the solution number 10 as1198911(119883) = 89298 (119905ℎ) and 119891
2(119883) = 95763 () The
corresponding decision variables are 119872MF = 89298 (119905ℎ)120601119861= 32119882
1= 75127 (119905ℎ) and119882
2= 16296 (119905ℎ)
5 Conclusions
Promoted by the requirements of engineering optimizationin complex practical processes of grinding and classificationwe proposed a hybrid multiobjective differential evolutionalgorithm with a few beneficial features integrated Firstly anarchiving mechanism for infeasible solutions is establishedwith functional partitioned multi-population which aims todirect the population to approach or land in the feasibleregion from different directions during evolution Secondlywe propose an infeasible constraint violations function toselect infeasible population with better performance sothat they are allowed to be saved and to participate inthe subsequent evolution Thirdly a nondominating rankingstrategy is designed to improve the crowding-distance sortingand return uniform distribution of Pareto solutions Finallythe simplex method is inserted in the differential evolutionprocess to purposefully enrich the diversity without excessivecomputation cost The advantage of the proposed algorithmis the exemption from constructing penalty function andthe preservation of meaningful infeasible solutions directlySimulation results on benchmarks indicate that the proposedalgorithm can converge quickly and effectively to the truePareto frontier with better distribution
Based on the investigated information about grinding cir-cuit process we established a multiobjective optimal modelwith equations frommechanism knowledge parameters rec-ognized by data regression and constraints of technicalrequirements The nonlinear multiobjective optimizationmodel is too complicated to be solved by traditional gradient-based algorithmsThe proposed hybrid differential evolutionalgorithm is applied and tested to achieve a Pareto solution
14 Journal of Applied Mathematics
set It is proven to be valuable for operation decision makingin the industrial process and showed superiority to the oper-ation carried out in the production In fact many operatingparameters in complex processes are highly coupled andconflicting with each other The optimal operation of theentire production process is very difficult to obtain bymanualcalculation let alone the fluctuation situation of processconditions The application case indicates that the proposedmethod has good performance and is helpful to inspirefurther research on evolutionary methods for engineeringoptimization
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgments
This work is supported by the National Natural ScienceFoundation of China (61374156 61273187 and 61134006)the National Key Technology Research and DevelopmentProgram of the Ministry of Science and Technology ofChina (2012BAK09B04) the Science Fund for CreativeResearch Groups of the National Natural Science Foundationof China (61321003) the Fund for Doctor Station of theMinistry of Education (20110162130011 and 20100162120019)and theHunanProvince Science andTechnology PlanProject(2012CK4018)
References
[1] K Mitra and R Gopinath ldquoMultiobjective optimization ofan industrial grinding operation using elitist nondominatedsorting genetic algorithmrdquo Chemical Engineering Science vol59 no 2 pp 385ndash396 2004
[2] G Yu T Y Chai and X C Luo ldquoMultiobjective productionplanning optimization using hybrid evolutionary algorithmsfor mineral processingrdquo IEEE Transactions on EvolutionaryComputation vol 15 no 4 pp 487ndash514 2011
[3] T Y Ma W H Gui Y L Wang and C H Yang ldquoDynamicoptimization control for grinding and classification processrdquoControl andDecision vol 27 no 2 pp 286ndash290 2012 (Chinese)
[4] F Y Edgeworth Mathematical Physics An Essay on the Appli-cation of Mathematics to the Moral Sciences C Kegan Paul ampCompany London UK 1881
[5] V Pareto Cours drsquoEconomie Politique F Rouge LausanneSwitzerland 1896
[6] X F Chen W H Gui Y L Wang and L Cen ldquoMulti-stepoptimal control of complex process a genetic programmingstrategy and its applicationrdquo Engineering Applications of Artifi-cial Intelligence vol 17 no 5 pp 491ndash500 2004
[7] C A C Coello ldquoEvolutionary multi-objective optimizationa historical view of the fieldrdquo IEEE Computational IntelligenceMagazine vol 1 no 1 pp 28ndash36 2006
[8] K Deb A Pratap S Agarwal and T Meyarivan ldquoA fastand elitist multiobjective genetic algorithm NSGA-IIrdquo IEEETransactions on Evolutionary Computation vol 6 no 2 pp 182ndash197 2002
[9] C M Fonesca and P J Fleming ldquoGenetic algorithms formulti objective optimization formulation discussion and gen-eralizationrdquo in Proceedings of the 5th International Conferenceon Genetic Algorithms pp 415ndash423 Morgan Kaufmann SanMateo Calif USA 1993
[10] S Sengupta S Das M Nasir A V Vasilakos and W PedryczldquoEnergy-efficient differentiated coverage of dynamic objectsusing an improved evolutionary multi-objective optimizationalgorithm with fuzzy-dominancerdquo in Proceedings of IEEECongress on Evolutionary Computation pp 1ndash8 IEEE PressBrisbane Australia 2012
[11] K V Price R M Storn and J A Lampinen DifferentialEvolution a Practical Approach to Global Optimization SpringerBerlin Germany 2005
[12] W Y Gong and Z H Cai ldquoAn improved multiobjective dif-ferential evolution based on Pareto-adaptive 120576-dominance andorthogonal designrdquo European Journal of Operational Researchvol 198 no 2 pp 576ndash601 2009
[13] X H Zeng W K Wong and S Y Leung ldquoAn operatorallocation optimization model for balancing control of thehybrid assembly lines using Pareto utility discrete differentialevolution algorithmrdquo Computers and Operations Research vol39 no 5 pp 1145ndash1159 2012
[14] E Zitzler and L Thiele ldquoMultiobjective evolutionary algo-rithms a comparative case study and the strength Paretoapproachrdquo IEEETransactions onEvolutionaryComputation vol3 no 4 pp 257ndash271 1999
[15] A Al-Ani A Alsukker and R N Khushaba ldquoFeature subsetselection using differential evolution and a wheel based searchstrategyrdquo Swarm and Evolutionary Computation vol 9 pp 15ndash26 2013
[16] G Y Li and M G Liu ldquoThe summary of differential evolutionalgorithm and its improvementsrdquo in Proceedings of the 3rdInternational Conference on Advanced Computer Theory andEngineering (ICACTE rsquo10) pp V3153ndashV3156 Chengdu ChinaAugust 2010
[17] H A Abbass ldquoThe self-adaptive Pareto differential evolutionalgorithmrdquo in Proceedings of the IEEE Congress on EvolutionaryComputation pp 831ndash836 IEEE Press Honolulu Hawaii USAMay 2002
[18] N K Madavan ldquoMultiobjective optimization using a Paretodifferential evolution approachrdquo in Proceedings of the IEEECongress on Evolutionary Computation pp 1145ndash1150 IEEEPress Honolulu Hawaii USA May 2002
[19] F Xue A C Sanderson and R J Graves ldquoPareto-based multi-objective differential evolutionrdquo in Proceedings of the IEEECongress on Evolutionary Computation pp 862ndash869 IEEEPress Piscataway NJ USA December 2003
[20] T Robic and B Filipi ldquoDEMO differential evolution formultiobjective optimizationrdquo in Evolutionary Multi-CriterionOptimization pp 520ndash533 Springer Berlin Germany 2005
[21] V L Huang A K Qin P N Suganthan and M F TasgetirenldquoMulti-objective optimization based on self-adaptive differen-tial evolution algorithmrdquo in Proceedings of the IEEE Congresson Evolutionary Computation (CEC rsquo07) pp 3601ndash3608 IEEEPress Singapore September 2007
[22] V L Huang S Z Zhao R Mallipeddi and P N SuganthanldquoMulti-objective optimization using self-adaptive differentialevolution algorithmrdquo in Proceedings of the IEEE Congress onEvolutionary Computation (CEC rsquo09) pp 190ndash194 IEEE PressTrondheim Norway May 2009
Journal of Applied Mathematics 15
[23] J A Adeyemo and F A O Otieno ldquoMulti-objective differentialevolution algorithm for solving engineering problemsrdquo Journalof Applied Sciences vol 9 no 20 pp 3652ndash3661 2009
[24] Y Wang and Z Cai ldquoCombining multiobjective optimizationwith differential evolution to solve constrained optimizationproblemsrdquo IEEE Transactions on Evolutionary Computationvol 16 no 1 pp 117ndash134 2012
[25] A Homaifar C X Qi and S H Lai ldquoConstrained optimizationvia genetic algorithmsrdquo Simulation vol 62 no 4 pp 242ndash2531994
[26] J A Joines and C R Houck ldquoOn the use of non-stationarypenalty functions to solve nonlinear constrained optimizationproblems with GArsquosrdquo in Proceedings of the 1st IEEE Conferenceon Evolutionary Computation pp 579ndash584 Orlando Fla USAJune 1994
[27] F Jimenez and J L Verdegay ldquoEvolutionary techniques forconstrained optimization problemsrdquo in Proceedings of the 7thEuropean Congress on Intelligent Techniques and Soft Computing(EUFIT rsquo99) Aachen Germany 1999
[28] C Sun J Zeng and J Pan ldquoAn improved vector particleswarm optimization for constrained optimization problemsrdquoInformation Sciences vol 181 no 6 pp 1153ndash1163 2011
[29] Z Michalewicz and M Schoenauer ldquoEvolutionary algorithmsfor constrained parameter optimization problemsrdquo Evolution-ary Computation vol 4 no 1 pp 1ndash32 1996
[30] C A C Coello and A D Christiansen ldquoMoses a multiob-jective optimization tool for engineering designrdquo EngineeringOptimization vol 31 no 1ndash3 pp 337ndash368 1999
[31] Y G Woldesenbet G G Yen and B G Tessema ldquoConstrainthandling in multiobjective evolutionary optimizationrdquo IEEETransactions on Evolutionary Computation vol 13 no 3 pp514ndash525 2009
[32] J A Nelder and R Mead ldquoA simplex method for functionminimizationrdquoThe Computer Journal vol 7 no 4 pp 308ndash3131965
[33] D A van Veldhuizen and G B Lamont ldquoMultiobjective evolu-tionary algorithms analyzing the state-of-the-artrdquo EvolutionaryComputation vol 8 no 2 pp 125ndash147 2000
[34] R Farmani D A Savic andG AWalters ldquoEvolutionarymulti-objective optimization in water distribution network designrdquoEngineering Optimization vol 37 no 2 pp 167ndash183 2005
[35] K Deb and T Goel Controlled Elitist Non-Dominated SortingGenetic Algorithms for Better Convergence Springer BerlinGermany 2001
[36] Y J Lai T Y Liu and C L Hwang ldquoTopsis for MODMrdquoEuropean Journal of Operational Research vol 76 no 3 pp486ndash500 1994
Submit your manuscripts athttpwwwhindawicom
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical Problems in Engineering
Hindawi Publishing Corporationhttpwwwhindawicom
Differential EquationsInternational Journal of
Volume 2014
Applied MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Probability and StatisticsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical PhysicsAdvances in
Complex AnalysisJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
OptimizationJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
CombinatoricsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Operations ResearchAdvances in
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Function Spaces
Abstract and Applied AnalysisHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of Mathematics and Mathematical Sciences
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Algebra
Discrete Dynamics in Nature and Society
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Decision SciencesAdvances in
Discrete MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014 Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Stochastic AnalysisInternational Journal of
6 Journal of Applied Mathematics
SM-DEMOPareto frontier
0 01 02 03 04 05 06 07 08 09 1f1
0
01
02
03
04
05
06
07
08
09
1f2
(a) ZDT2
0
02
04
06
08
1
minus08
minus06
minus04
minus02
f2
SM-DEMOPareto frontier
0 01 02 03 04 05 06 07 08 09 1f1
(b) ZDT3
SM-DEMOPareto frontier
0 01 02 03 04 05 06 07 08 09 1f1
0
01
02
03
04
05
06
07
08
09
1
f2
(c) ZDT4
SM-DEMO
0
01
02
03
04
05
06
07
08
09
1
Pareto frontier
02 03 04 05 06 07 08 09 1f1
f2
(d) ZDT6
Figure 2 SM-DEMO simulation curve
Assume that 119909lowast is the global optimal solution and 1199101is
the closest one to 119909lowast If the infeasible population 1199101is not
excluded by the evolution algorithm it is permitted to exploreboundary regions from new directions where the optimumis likely to be found
31 General Idea of the Proposed Algorithm Researchershave gradually realized the merit of infeasible solutions insearching for the global optimum in the feasible region Someinfeasible solutions with better performance are allowed to besaved Farmani et al [34] formulated a method to ensure thatinfeasible solutions with a slight violation become feasible inevolution Based on the constraints processing approach ofmultiobjective optimization problems the proposed hybridDE algorithm avoids constructing penalty function anddeleting meaningful infeasible solutions directly
Here the proposed algorithm will produce multiplegroups of functional partitions which include an evolution-ary population 119875119892 of size 119873119901 an intermediate population119872119892 to save feasible individuals an intermediate population119878119892 to save infeasible individuals a population 119875119891 to save theoptimal feasible solution found in the search process and apopulation 119875119888 to save the optimal infeasible solution Therelationship of multi-population is shown in Figure 4
With the description of (1) equality constraints are alwaystransformed into inequality constraints as |ℎ
119895(119883)| minus 120575 le 0
where 119895 = 119901 + 1 119902 and 120575 is a positive tolerance valueTo evaluate the infeasible solution the degree of constraintviolation of individual 119883 on the 119895th constraint is calculatedas follows
119881119894(119883) =
max 0 119892119894(119883) (119894 = 1 2 119901)
max 0 10038161003816100381610038161003816ℎ119895 (119883)10038161003816100381610038161003816minus 120575 (119895 = 119901 + 1 119902)
(8)
Journal of Applied Mathematics 7
Feasible solutionspace Ω
Search space S
x2
x1y1
y4
y3
y2 x3
x4
xlowast
Infeasible solutionspace Z
Figure 3 Distribution diagram of search space
Evolutionarypopulation
Pg
Intermediatepopulation
Mg
Optimal Randomly selected
Intermediatepopulation
SgInfeasible
Feasibleindividuals
individuals
Randomlyselected
Evolution
g = g + 1
g = g + 1
g = g + 1Randomly selected
feasiblesolution Pf
Optimal infeasible
solution Pc
Figure 4The relationship diagram ofmultipopulation demonstrat-ing
The final constraint violation of each individual in thepopulation can be obtained by calculating the mean of thenormalized constraint violations
In order to take advantage of the infeasible solutionswith better performance we proposed the following adaptivedifferential mutation operator to generate individual varia-tion learning from the mutation operators in DErand-to-best1bin according to rules defined by Price et al [11]Considering each individual vector 119909119866
119894 a mutant individual
119909119866+1
119894is defined by
119909119866+1
119894
= 119909119866
119894+ 1198651sdot (119909119866
1198911minus 119909119866
1199031)+ 1198652sdot (119909119866
1198912minus 119909119866
1199032) 119877
119862ge rand ( )
119909119866
119894+ 1198651sdot (119910119866
119894minus 119909119866
1199031)+ 1198652sdot (119909119866
1198911minus 119909119866
1199032) 119877
119862lt rand ( )
(9)
where 1199031and 1199032represent different integers and also different
from 119894 randomly generated over [1119873119901] 119865 is the scalingfactor 119909119866
119891119894(119894 = 1 2 119899) is randomly generated from 119875119891
119910119866
119894(119894 = 1 2 119899) is randomly generated from 119875119888 and 119877
119862is
the mutation factor as follows
119877119862= 1198771198880sdot120574119866+ const
120574119866minus1 + const (10)
where 1198771198880is the initial value of the variability factor const is
a small constant to ensure that the fractional is meaningfuland 120574119866 is defined as follows
120574119866=the number of infeasible solutions in 119875119892
119873119901 (11)
32 Framework of the Proposed Algorithm The proposedalgorithm is described as follows
Step 1 (initialization) Generate the population 119875119892 119875119891 and119875119888 of size 119873119901 119873119875
1 and 119873119875
2 Set the value of 119862
119877(crossover
probability)119866max (the number of function evaluations)119866SM(the iterative number of evolution by NM simplex method)119892 = 1 (the current generation number) and positive controlparameter for scaling the difference vectors 119865
1 1198652 Randomly
generate the parent population 119875119892 from the decision spaceSet the 119875119891 and 119875119888 and let the intermediate populations 119878119892and119872119892 be empty
Step 2 (DE reproduction) By (3) and (9) for mutationcrossover and selection an offspring 119878119892 is created Judge theconstraints of all individuals in 119875119892 In accordance with (8)we first calculate constraint violation degree 119881
119894(119883) of all of
the individuals If 119881119894(119883) = 0 the solution is feasible and
preserved to the intermediate set 119872119892 if 119881119894(119883) gt 0 the
solution is infeasible and preserved to the intermediate set 119878119892
Step 3 (simplex method) Apply NM simplex method oper-ator to the present population if 119892 ge 119866max2 Updatethe present population 119872119892 when the number of iterationsexceeds maximum iterations
Step 4 (119875119891 construction) Rank chromosomes in 119872119892 basedon (5) and generate the elitist population 119875119891 (the size is119873119901)from the ranked population119872119892
Step 5 (119875119888 construction) Add the chromosomes in 119878119892 withslight constraint violations to the 119875119888
Step 6 (mixing the population) Combine 119878119892with the existingparent population to form a new set119872119892 Remove the dupli-cate individuals in119872119892 and the existing parent population
Step 7 (evolution) Randomly choose chromosomes from119875119888 119875119892 and 119875119891 Use the adaptive differential mutation anduniform discrete crossover to obtain the offspring population119875119892 + 1
Step 8 (termination) If the stopping criterion ismet stop andoutput the best solution else go to Step 2
33 Experimental Study In this section we choose threeproblems CTP TNK and BNH as shown in Table 3 to testthe proposed method and compare the method with thecurrent CNSGA-II [35]
8 Journal of Applied Mathematics
Table 3 Test functions
Testfunction
Objective functionmin119865(119883) = min[119891
1(119883) 119891
2(119883)]
Constraints Range ofvariable
CTP
1198911(119883) = 119909
1
1198912(119883) = exp(minus
1198911(119883)
119888 (119883))
times41 +
5
sum
119894=2
[119909119894
2minus 10 cos (2120587119909
119894)]
1198921(119883) = cos (120579) (119891
2(119883) minus 119890) minus sin (120579) 119891
1(119883)
1198922(119883) = 119886
1003816100381610038161003816sin 119887120587 [sin (120579) (1198912 (119883) minus 119890)+ cos (120579) 119891
1(119883)]119888
10038161003816100381610038161003816
119889
1198921(119883) ge 119892
2(119883)
0 le 1199091le 1
minus5 le 119909119894le 5
119894 = 2 3 4 5
BNH1198911(119883) = 4119909
1
2+ 41199092
2
1198912(119883) = (119909
1minus 5)2+ (1199092minus 5)2
1198921(119883) = (119909
1minus 5)2+ 1199092
2minus 25
1198922(119883) = minus(119909
1minus 8)2+ (1199092+ 3) + 77
1198921(119883) le 0 119892
2(119883) le 0
0 le 1199091le 5
0 le 1199092le 3
TNK1198911(119883) = 119909
1
1198912(119883) = 119909
2
1198921(119883) = minus119909
1
2minus 1199092
2+ 1
+01 cos(16 arctan(11990911199092
))
1198922(119883) = (119909
1minus 05)
2+ (1199092minus 05)
1198921(119883) le 0 119892
2(119883) minus 05 le 0
0 le 119909119894le 120587
119894 = 1 2
20
15
10
5
00
Feasible solutionInfeasible solution
02 04 06 08 10f1
f2
Pareto-optimal frontier
(a) CTP1
20
15
10
05
0002 04 06 08 10
Feasible solutionInfeasible solution
Pareto-optimal frontier
f1
f2
(b) CTP2
Figure 5 CTP solution space
For CTP problem there are the six parameters 120579 119886 119887 119888119889 and 119890 that must be chosen in a way so that a portion ofthe unconstrained Pareto-optimal region is infeasible Eachconstraint is an implicit non-linear function of decisionvariables Thus it may be difficult to find a number ofsolutions on a non-linear constraint boundary We take twosets of values for six parameters in CTP problem which aredetermined as (1) CTP1 120579 = 01120587 119886 = 40 119887 = 05 119888 = 1119889 = 2 and 119890 = minus2 (2) CTP2 120579 = minus02120587 119886 = 02 119887 = 10119888 = 1 119889 = 6 and 119890 = 1 The Pareto frontiers the feasiblesolution spaces and the infeasible solution spaces are shownin Figure 5
The parameters are initialized as followsThe size of pop-ulation 119875119892 is119873119901 = 200 size of 119875119891 is119873119875
1= 150 size of 119875119888 is
1198731198752= 10 scaling factors 119865
1and 119865
2are randomly generated
within [05 1] cross rate is 119862119877= 06 maximum evolution
generation is 119866max = 300 and number of SM evolutioniterations is 119866SM = 100 All of the proposed algorithmsand CNSGA-II are real coded with equal population size andequal maximum evolution generation Each algorithm runs20 times independently for each test function
Figure 6 shows the result of a random running of theproposed algorithm and NSGA-II the smooth curve ldquomdashrdquorepresents the Pareto frontier and ldquoXrdquo stands for the solutionachieved by the proposed algorithm or NSGA-II
It is obvious that the proposed algorithm returns abetter approximation to the true Pareto-optimal frontie anda distribution of higher uniformity We also evaluated the
Journal of Applied Mathematics 9
40
35
30
25
20
15
10
050
Pareto frontierProposed algorithm
02 0301 04 05 07 0906 08 10
f2
f1
(a) CTP1 calculated by the proposed algorithm
CNSGA-IIPareto frontier
40
35
30
25
20
15
10
05
f2
f1
0 02 0301 04 05 07 0906 08 10
(b) CTP1 calculated by CNSGA-II
1211100908070605040302
f2
Pareto frontierProposed algorithm
f1
0 02 0301 04 05 07 0906 08 10
(c) CTP2 calculated by the proposed algorithm
CNSGA-IIPareto frontier
1211100908070605040302
f2
f1
0 02 0301 04 05 07 0906 08 10
(d) CTP2 calculated by CNSGA-II
50
40
30
20
10
00 50 100 150
Pareto frontierProposed algorithm
f2
f1
(e) BNH calculated by the proposed algorithm
Pareto frontier
50
40
30
20
10
00 50 100 150
f2
f1
CNSGA-II
(f) BNH calculated by CNSGA-II
Figure 6 Continued
10 Journal of Applied Mathematics
15
10
05
00 05 10 15
Pareto frontierProposed algorithm
f2
f1
(g) TNK calculated by the proposed algorithm
Pareto frontier
15
10
05
00 05 10 15
f2
f1
CNSGA-II
(h) TNK calculated by CNSGA-II
Figure 6 The comparison chart of Pareto frontier
Table 4 The comparison of performance
Test function Algorithm 120574 SP
CTP1 CNSGA-Π 0021317 plusmn 0000323 0873321 plusmn 008725
The proposed 0009836 plusmn 0000410 0567933 plusmn 001845
CTP2 CNSGA-Π 0011120 plusmn 0000753 078314 plusmn 002843
The proposed 0007013 plusmn 0000554 0296543 plusmn 000453
BNH CNSGA-Π 0014947 plusmn 0000632 0336941 plusmn 000917
The proposed 0013766 plusmn 0000043 0209321 plusmn 000561
TNK CNSGA-Π 0013235 plusmn 0000740 0464542 plusmn 000730
The proposed 0006435 plusmn 0000017 0224560 plusmn 000159
Feed water Water
Feed ore Ball mill OverflowMixing
PumpClassifierwater
additionFirst
classifierrecycle First-
overflow Spiral classifier 2Second-stage
to flotationSpiral classifier 1 Second classifier recycle(rough concentrate)
addition
pump
stageoverflow
Figure 7 Flow diagram of the grinding and classification process
algorithms based on the two criterions 120574 and SP as shownin Table 4 It can be observed from the data in Table 4that the proposed algorithm performs significantly betterthan the classical CNSGA-II algorithm in convergence anddistribution uniformityThe simulation results show that thisalgorithm can accurately converge to global Pareto solutionsand can maintain diversity of population
4 Optimization of Grinding andClassification Process
41 Bauxite Grinding and Classification Process Thegrindingand classification process is the key preparation for thebauxite mineral processing Here we consider a bauxitegrinding process in a certain mineral company with singlegrinding and two-stage classification as shown in Figure 7
The process consists in a grinding ball mill and twospiral classifiers First classifier recycle will be put back tothe ball mill for regrinding and the first-stage overflowwill be put into second spiral classifier after being mixedwith water the second classifier recycle will be prepared forBayer production as the rough concentrate and the second-stage overflow will be sent to the next flotation processThe production objectives are composed of the productionyields technically represented by the solid flow of feed oresince the process is nonstorable and the mineral processingquality represented by percentage of the small-size fractionsof mineral particles in the second-stage overflows
Journal of Applied Mathematics 11
Table 5 Notations for the model of the grinding and classificationprocess
Notation Description
119901119894
Particle percentage of the 119894th size fraction inthe ball mill overflow
119891119895
Particle percentage of the 119894th size fraction inthe Feed ore
119862 Rate of the first classifier recycle119864119886119894 The efficiency of the first spiral classifier
120591 Themean residence time1198751 The internal concentration in ball mill119872MF The solid flow of feed ore1198821 The water addition of the first classifier recycle
1198822 The classifier water addition
119887119894119895 The breakage distribution function119878119894 The breakage rate function119889119894 The particle with the 119894th size
120572119888119894
The particle percentage of the 119894th size in thefirst classifier overflow
119889119900 The unit size of the particle1198752 First-stage overflow120601119861 Ball filling rate
119860119888 119861119888
Parameters of the first-stage classifier overflowsize fraction distribution
119864119886119894 The efficiency of the first spiral classifier
11988950119888
The particle size fraction after correctionseparation
119898 The separation accuracy119896 The intermix index
1198641198861015840
min 1198891015840
50119888 1198981015840 1198961015840 The corresponding key parameters to theefficiency of the second spiral classifier
119886 120572 120583 ΛFour key parameters to control the breakagerate function
119889min 119889max Theminimum and maximum particle sizes119860MF 119861MF Parameters of feed ore size fraction distribution
119865(119894)The cumulative particle percentage less thanthe 119894th size fraction in feed ore
1205721198881015840
119894
The particle percentage of the 119894th size fractionin the second classifier overflow
1198641198861015840
119894 The efficiency of the second spiral classifier
42 Predictive Model of the Grinding and Classification Pro-cess Here we establish the mathematical predictive modelof each unit process in the bauxite grinding and classificationprocess The notations of the indexes decision variables andparameters are listed in Table 5 These notations will be usedfor the model of the grinding and classification process
421 Ball Mill Circuit Model Here 119901119894is the particle per-
centage of 119894th size fraction in the ball mill overflow 119891119895is the
particle percentage of 119894th size fraction in feed ore rate of thefirst classifier recycle 119862 is known and 119864119886
119894is the efficiency of
the first spiral classifier According to a technical report offield investigation and study we have that
119901119894(1 + 119862) =
119889119894119894119891119894+ sum119894minus1
119895=1119894gt1119889119894119895[119864119886119894119901119895(1 + 119862) + 119891
119895]
1 minus 119889119894119894119864119886119894
119889119894119895=
119890119895 119894 = 119895
119894minus1
sum
119896=119895
119888119894119896119888119895119896(119890119896minus 119890119894) 119894 gt 119895
119888119894119895=
minus
119895minus1
sum
119896=119894
119888119894119896119888119895119896 119894 lt 119895
1 119894 = 119895
1
119878119894minus 119878119895
119894minus1
sum
119896=119895
119878119896119887119894119896119888119896119895 119894 gt 119895
119890119895=
1
(1 + 05 sdot 120591119878119895) (1 + 025 sdot 120591119878
119895)2 120591 =
81198751
119872MF (12)
where 120591 is the mean residence time of minerals 1198751is the
internal concentration in ball mill and
1198751=
119872MF (1 + 119862)
119872MF (1 + 119862) +1198821 + 03 (1198821 +1198822) (13)
where119872MF (119905ℎ) is the solid flow of feed ore1198821is the water
addition of the first classifier recycle and1198822is the classifier
water addition 119887119894119895is the breakage distribution function 119878
119894
is the breakage rate function and it satisfied the followingequation
119878119894=
(119886(119889119894119889119900)120572
)
(1 + (119889119894120583)Λ
)
(14)
where 119889119894is the particle with the 119894th size 119889
119900it is a unit when
per millimeter is a unit 119889119900= 1 119889
119894= 119894 (mm) and 119886 120572 120583
and Λ are four key parameters to control the breakage ratefunction
In a concrete grinding and classification process the ballmill size is fixed and the speed of ball mill is constantThrough data acquisition and testing of grinding and clas-sification steady-state loop the regression model between 119886120572 120583 Λ and condition variables size fraction distributioncan be established The input variables are ball filling rate120601119861 solid flow of feed ore 119872MF water addition of the first
classifier recycle1198821 and parameters of feed ore size fraction
distribution 119860MF 119861MF The regression model is
[119886 120572 120583 Λ]119879
=[[
[
11990911
sdot sdot sdot 1199091119895
d
1199091198941
sdot sdot sdot 119909119894119895
]]
]
sdot [1198821 120601119861 119860MF 119861MF 119872MF]119879
(15)
The value of 119909119894119895(119894 = 1 2 3 4 119895 = 1 2 5) can be
obtained by the experimental data regression 119860MF 119861MF can
12 Journal of Applied Mathematics
be obtained from feed ore size fraction distribution and 119865(119894)is the cumulative particle percentage less than the 119894th sizefraction in feed ore and it is represented as follows
119865 (119894) = 1 minus exp (minus119860MF119889119861MF119894
) (16)
422 Spiral Classifier Model 120572119888119894is the particle percentage
of the 119894th size fraction in the first classifier overflow and 119901119894
is the particle percentage of the 119894th size fraction in ball milloverflow The spiral classifier model is as follows
120572119888119894=
119901119894times (1 minus 119864119886
119894)
sum119894=1
(119901119894times (1 minus 119864119886
119894))times 100 (17)
where 119864119886119894is the efficiency of the first spiral classifier and the
mechanism formula of 119864119886119894is shown as follows
119864119886119894= 1 minus exp [minus0693(
119889119894minus 119889min
11988950119888
minus 119889min)
119898
]
+ 119864119886min sdot [1 minus (119889119894minus 119889min119889max
)]
119896
(18)
where 119889119894is the particle with the 119894th size 119889min and 119889max
represent maximum and minimum particle sizes 11988950119888
isthe particle size fraction after correction separation 119898 isseparation accuracy and 119896 is intermix index
Through data acquisition and testing of grinding andclassification steady-state loop the regressionmodel betweenclassification parameters and condition variables size frac-tion distribution can be established The input variablesinclude the solid flow of feed ore 119872MF the classifier wateraddition 119882
2 and the parameters of ball mill overflow size
fraction distribution 119860MF 119861MF The regression model isshown as follows
[119864119886min 11988950119888
119898 119896]119879
=[[
[
11991011
sdot sdot sdot 1199101119895
d
1199101198941
sdot sdot sdot 119910119894119895
]]
]
sdot [119872MF 1198822119860MP 119861MP]
119879
(19)
where the value of 119910119894119895(119894 = 1 2 3 4 119895 = 1 2 3 4) can be
obtained by data regressionThe first-stage overflow 119875
2calculation formula is as
follows
1198752=
119872MF(119872MF +1198821 +1198822)
(20)
Similarly we can get the second spiral classifier model asfollows
1205721198881015840
119894=
120572119888119894times (1 minus 119864119886
1015840
119894)
sum119894(120572119888119894times (1 minus 1198641198861015840))
times 100 (21)
where 1205721198881015840119894is the particle percentage of the 119894th size fraction in
the second classifier overflow 120572119888119894is the particle percentage
of the 119894th size fraction in the first classifier overflow and
1198641198861015840
119894is the efficiency of the second spiral classifier The spiral
classifier model is as follows
1198641198861015840
119894= 1 minus exp[
[
minus0693(1198891015840
119894minus 1198891015840
min1198891015840
50119888minus 1198891015840
min)
1198981015840
]
]
+ 1198641198861015840
min sdot [1 minus (1198891015840
119894minus 1198891015840
min1198891015840max
)]
1198961015840
(22)
where 1198641198861015840min 1198891015840
50119888 1198981015840 and 119896
1015840 are key parameters to theefficiency of the second spiral classifier Through data acqui-sition and testing of grinding and classification steady-stateloop the regression model between classification parametersand condition variables size fraction distribution can beestablished The input variables include solid flow of feedore119872MF and parameters of the first-stage classifier overflowsize fraction distribution 119860
119888 119861119888 which are solved by similar
equation to (20) The regression model is shown as follows
[1198641198861015840
min 1198891015840
5011988811989810158401198961015840]119879
=
[[[
[
1199101015840
11sdot sdot sdot 1199101015840
1119895
d
1199101015840
1198941sdot sdot sdot 119910
1015840
119894119895
]]]
]
sdot [119872MF 119860119888119861119888]119879
(23)
where the value of1199101015840119894119895(119894 = 1 2 3 4 119895 = 1 2 3) can be obtained
by experimental data regression
43 Optimization Model of Grinding and Classification Pro-cess Two objective functions in the process are identifiedone is tomaximize output119891
1(119883) and the other is tomaximize
the small-size fractions (less than 0075mm fractions) in thesecond-stage overflow 119891
2(119883) It is also necessary to ensure
that the grinding productmeets all of other technical require-ments and the least disturbance in the following flotationcircuit As the constraints the feed load of the grinding circuit119872MF the steel ball filling rate 120601
119861 the first and the second
overflows 1198751and 119875
2 and the particle percentage of fine size
fraction in the first and the second classifier overflows 120572119888minus0075
and 1205721198881015840minus0075
should be within the user specified boundsThe operation variables are the solid flow of feed ore
119872MF water addition of the first classifier recycle 1198821 ball
filling rate 120601119861 and water addition of the second classifier119882
2
Based on all of the above grinding and classification processmultiobjective optimization model is as follows
max 119865 = max [1198911(119883) 119891
2(119883)]
1198911(119883) = 119872MF
1198912(119883) = 120572119888
1015840
minus0075
= 119891 (119872MF11988211198822 120601119861)
st 119872MFmin le 119872MF le 119872MFmax120601119861min le 120601119861 le 120601119861max1198751min le 1198751 le 1198751max1198752min le 1198752 le 1198752max120572119888minus0075
ge 120572119888min
1205721198881015840
minus0075ge 1205721198881015840
min
(24)
Journal of Applied Mathematics 13
Table 6 The optimization results calculated by the proposed algo-rithm
Number 1198911(119883) (th) 119891
2(119883) ()
1 92972 908142 91810 919003 92360 909234 90620 934605 91310 924946 90170 938007 89390 952008 89480 949329 91530 9240010 89298 9576311 89824 9454912 89800 9439513 89710 9461814 91170 9280015 91880 9184016 92190 9128117 89870 9409018 91000 9328519 91960 9152020 91124 93221
85
87
89
91
93
95
97
74000 79000 84000 94000
Optimization resultsThe raw data
f1(x)
f2(x)
()
89000
Figure 8 The comparison chart between optimization results andindustrial data
In (24) 119891(119872MF11988211198822 120601119861) implicates the model ofgrinding and classification represented by (12)ndash(23) Theproposed algorithm is applied to solve the problem and theoptimization results are shown in Table 6
With the practical process data from a grinding circuitof a mineral plant the simulation of this hybrid intelligentmethod adopted the same parameters on the variation infresh slurry feed velocity density particle size distributionand cyclone feed operating configurations
The comparison of production data and optimizationresults in Table 6 is shown in Figure 8 where ldquoXrdquo repre-sents the proposed algorithm optimization results and ldquoIrdquorepresents the original data collected from the field withoutoptimization of raw data According to the objectives thedata point closer to the upper right edge is more beneficialObviously the proposed optimization result is far betterthan the original data indicating the effectiveness of theoptimization approach
44 TOPSIS Method for Solution Selection The resolution ofamultiobjective optimization problemdoes not endwhen thePareto-optimal set is found In practical operational prob-lems a single solution must be selected TOPSIS [36] is auseful technique in dealing with multiattribute or multicrite-ria decision-making (MADMMCDM) problems in the realworld The standard TOPSIS method attempts to choosealternatives that simultaneously have the shortest distancefrom the positive-ideal solution and the farthest distancefrom the negative-ideal solution According to the TOPSISmethod the relative closeness coefficient is calculated andthe best solution in Table 6 is the solution number 10 as1198911(119883) = 89298 (119905ℎ) and 119891
2(119883) = 95763 () The
corresponding decision variables are 119872MF = 89298 (119905ℎ)120601119861= 32119882
1= 75127 (119905ℎ) and119882
2= 16296 (119905ℎ)
5 Conclusions
Promoted by the requirements of engineering optimizationin complex practical processes of grinding and classificationwe proposed a hybrid multiobjective differential evolutionalgorithm with a few beneficial features integrated Firstly anarchiving mechanism for infeasible solutions is establishedwith functional partitioned multi-population which aims todirect the population to approach or land in the feasibleregion from different directions during evolution Secondlywe propose an infeasible constraint violations function toselect infeasible population with better performance sothat they are allowed to be saved and to participate inthe subsequent evolution Thirdly a nondominating rankingstrategy is designed to improve the crowding-distance sortingand return uniform distribution of Pareto solutions Finallythe simplex method is inserted in the differential evolutionprocess to purposefully enrich the diversity without excessivecomputation cost The advantage of the proposed algorithmis the exemption from constructing penalty function andthe preservation of meaningful infeasible solutions directlySimulation results on benchmarks indicate that the proposedalgorithm can converge quickly and effectively to the truePareto frontier with better distribution
Based on the investigated information about grinding cir-cuit process we established a multiobjective optimal modelwith equations frommechanism knowledge parameters rec-ognized by data regression and constraints of technicalrequirements The nonlinear multiobjective optimizationmodel is too complicated to be solved by traditional gradient-based algorithmsThe proposed hybrid differential evolutionalgorithm is applied and tested to achieve a Pareto solution
14 Journal of Applied Mathematics
set It is proven to be valuable for operation decision makingin the industrial process and showed superiority to the oper-ation carried out in the production In fact many operatingparameters in complex processes are highly coupled andconflicting with each other The optimal operation of theentire production process is very difficult to obtain bymanualcalculation let alone the fluctuation situation of processconditions The application case indicates that the proposedmethod has good performance and is helpful to inspirefurther research on evolutionary methods for engineeringoptimization
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgments
This work is supported by the National Natural ScienceFoundation of China (61374156 61273187 and 61134006)the National Key Technology Research and DevelopmentProgram of the Ministry of Science and Technology ofChina (2012BAK09B04) the Science Fund for CreativeResearch Groups of the National Natural Science Foundationof China (61321003) the Fund for Doctor Station of theMinistry of Education (20110162130011 and 20100162120019)and theHunanProvince Science andTechnology PlanProject(2012CK4018)
References
[1] K Mitra and R Gopinath ldquoMultiobjective optimization ofan industrial grinding operation using elitist nondominatedsorting genetic algorithmrdquo Chemical Engineering Science vol59 no 2 pp 385ndash396 2004
[2] G Yu T Y Chai and X C Luo ldquoMultiobjective productionplanning optimization using hybrid evolutionary algorithmsfor mineral processingrdquo IEEE Transactions on EvolutionaryComputation vol 15 no 4 pp 487ndash514 2011
[3] T Y Ma W H Gui Y L Wang and C H Yang ldquoDynamicoptimization control for grinding and classification processrdquoControl andDecision vol 27 no 2 pp 286ndash290 2012 (Chinese)
[4] F Y Edgeworth Mathematical Physics An Essay on the Appli-cation of Mathematics to the Moral Sciences C Kegan Paul ampCompany London UK 1881
[5] V Pareto Cours drsquoEconomie Politique F Rouge LausanneSwitzerland 1896
[6] X F Chen W H Gui Y L Wang and L Cen ldquoMulti-stepoptimal control of complex process a genetic programmingstrategy and its applicationrdquo Engineering Applications of Artifi-cial Intelligence vol 17 no 5 pp 491ndash500 2004
[7] C A C Coello ldquoEvolutionary multi-objective optimizationa historical view of the fieldrdquo IEEE Computational IntelligenceMagazine vol 1 no 1 pp 28ndash36 2006
[8] K Deb A Pratap S Agarwal and T Meyarivan ldquoA fastand elitist multiobjective genetic algorithm NSGA-IIrdquo IEEETransactions on Evolutionary Computation vol 6 no 2 pp 182ndash197 2002
[9] C M Fonesca and P J Fleming ldquoGenetic algorithms formulti objective optimization formulation discussion and gen-eralizationrdquo in Proceedings of the 5th International Conferenceon Genetic Algorithms pp 415ndash423 Morgan Kaufmann SanMateo Calif USA 1993
[10] S Sengupta S Das M Nasir A V Vasilakos and W PedryczldquoEnergy-efficient differentiated coverage of dynamic objectsusing an improved evolutionary multi-objective optimizationalgorithm with fuzzy-dominancerdquo in Proceedings of IEEECongress on Evolutionary Computation pp 1ndash8 IEEE PressBrisbane Australia 2012
[11] K V Price R M Storn and J A Lampinen DifferentialEvolution a Practical Approach to Global Optimization SpringerBerlin Germany 2005
[12] W Y Gong and Z H Cai ldquoAn improved multiobjective dif-ferential evolution based on Pareto-adaptive 120576-dominance andorthogonal designrdquo European Journal of Operational Researchvol 198 no 2 pp 576ndash601 2009
[13] X H Zeng W K Wong and S Y Leung ldquoAn operatorallocation optimization model for balancing control of thehybrid assembly lines using Pareto utility discrete differentialevolution algorithmrdquo Computers and Operations Research vol39 no 5 pp 1145ndash1159 2012
[14] E Zitzler and L Thiele ldquoMultiobjective evolutionary algo-rithms a comparative case study and the strength Paretoapproachrdquo IEEETransactions onEvolutionaryComputation vol3 no 4 pp 257ndash271 1999
[15] A Al-Ani A Alsukker and R N Khushaba ldquoFeature subsetselection using differential evolution and a wheel based searchstrategyrdquo Swarm and Evolutionary Computation vol 9 pp 15ndash26 2013
[16] G Y Li and M G Liu ldquoThe summary of differential evolutionalgorithm and its improvementsrdquo in Proceedings of the 3rdInternational Conference on Advanced Computer Theory andEngineering (ICACTE rsquo10) pp V3153ndashV3156 Chengdu ChinaAugust 2010
[17] H A Abbass ldquoThe self-adaptive Pareto differential evolutionalgorithmrdquo in Proceedings of the IEEE Congress on EvolutionaryComputation pp 831ndash836 IEEE Press Honolulu Hawaii USAMay 2002
[18] N K Madavan ldquoMultiobjective optimization using a Paretodifferential evolution approachrdquo in Proceedings of the IEEECongress on Evolutionary Computation pp 1145ndash1150 IEEEPress Honolulu Hawaii USA May 2002
[19] F Xue A C Sanderson and R J Graves ldquoPareto-based multi-objective differential evolutionrdquo in Proceedings of the IEEECongress on Evolutionary Computation pp 862ndash869 IEEEPress Piscataway NJ USA December 2003
[20] T Robic and B Filipi ldquoDEMO differential evolution formultiobjective optimizationrdquo in Evolutionary Multi-CriterionOptimization pp 520ndash533 Springer Berlin Germany 2005
[21] V L Huang A K Qin P N Suganthan and M F TasgetirenldquoMulti-objective optimization based on self-adaptive differen-tial evolution algorithmrdquo in Proceedings of the IEEE Congresson Evolutionary Computation (CEC rsquo07) pp 3601ndash3608 IEEEPress Singapore September 2007
[22] V L Huang S Z Zhao R Mallipeddi and P N SuganthanldquoMulti-objective optimization using self-adaptive differentialevolution algorithmrdquo in Proceedings of the IEEE Congress onEvolutionary Computation (CEC rsquo09) pp 190ndash194 IEEE PressTrondheim Norway May 2009
Journal of Applied Mathematics 15
[23] J A Adeyemo and F A O Otieno ldquoMulti-objective differentialevolution algorithm for solving engineering problemsrdquo Journalof Applied Sciences vol 9 no 20 pp 3652ndash3661 2009
[24] Y Wang and Z Cai ldquoCombining multiobjective optimizationwith differential evolution to solve constrained optimizationproblemsrdquo IEEE Transactions on Evolutionary Computationvol 16 no 1 pp 117ndash134 2012
[25] A Homaifar C X Qi and S H Lai ldquoConstrained optimizationvia genetic algorithmsrdquo Simulation vol 62 no 4 pp 242ndash2531994
[26] J A Joines and C R Houck ldquoOn the use of non-stationarypenalty functions to solve nonlinear constrained optimizationproblems with GArsquosrdquo in Proceedings of the 1st IEEE Conferenceon Evolutionary Computation pp 579ndash584 Orlando Fla USAJune 1994
[27] F Jimenez and J L Verdegay ldquoEvolutionary techniques forconstrained optimization problemsrdquo in Proceedings of the 7thEuropean Congress on Intelligent Techniques and Soft Computing(EUFIT rsquo99) Aachen Germany 1999
[28] C Sun J Zeng and J Pan ldquoAn improved vector particleswarm optimization for constrained optimization problemsrdquoInformation Sciences vol 181 no 6 pp 1153ndash1163 2011
[29] Z Michalewicz and M Schoenauer ldquoEvolutionary algorithmsfor constrained parameter optimization problemsrdquo Evolution-ary Computation vol 4 no 1 pp 1ndash32 1996
[30] C A C Coello and A D Christiansen ldquoMoses a multiob-jective optimization tool for engineering designrdquo EngineeringOptimization vol 31 no 1ndash3 pp 337ndash368 1999
[31] Y G Woldesenbet G G Yen and B G Tessema ldquoConstrainthandling in multiobjective evolutionary optimizationrdquo IEEETransactions on Evolutionary Computation vol 13 no 3 pp514ndash525 2009
[32] J A Nelder and R Mead ldquoA simplex method for functionminimizationrdquoThe Computer Journal vol 7 no 4 pp 308ndash3131965
[33] D A van Veldhuizen and G B Lamont ldquoMultiobjective evolu-tionary algorithms analyzing the state-of-the-artrdquo EvolutionaryComputation vol 8 no 2 pp 125ndash147 2000
[34] R Farmani D A Savic andG AWalters ldquoEvolutionarymulti-objective optimization in water distribution network designrdquoEngineering Optimization vol 37 no 2 pp 167ndash183 2005
[35] K Deb and T Goel Controlled Elitist Non-Dominated SortingGenetic Algorithms for Better Convergence Springer BerlinGermany 2001
[36] Y J Lai T Y Liu and C L Hwang ldquoTopsis for MODMrdquoEuropean Journal of Operational Research vol 76 no 3 pp486ndash500 1994
Submit your manuscripts athttpwwwhindawicom
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical Problems in Engineering
Hindawi Publishing Corporationhttpwwwhindawicom
Differential EquationsInternational Journal of
Volume 2014
Applied MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Probability and StatisticsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical PhysicsAdvances in
Complex AnalysisJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
OptimizationJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
CombinatoricsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Operations ResearchAdvances in
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Function Spaces
Abstract and Applied AnalysisHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of Mathematics and Mathematical Sciences
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Algebra
Discrete Dynamics in Nature and Society
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Decision SciencesAdvances in
Discrete MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014 Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Stochastic AnalysisInternational Journal of
Journal of Applied Mathematics 7
Feasible solutionspace Ω
Search space S
x2
x1y1
y4
y3
y2 x3
x4
xlowast
Infeasible solutionspace Z
Figure 3 Distribution diagram of search space
Evolutionarypopulation
Pg
Intermediatepopulation
Mg
Optimal Randomly selected
Intermediatepopulation
SgInfeasible
Feasibleindividuals
individuals
Randomlyselected
Evolution
g = g + 1
g = g + 1
g = g + 1Randomly selected
feasiblesolution Pf
Optimal infeasible
solution Pc
Figure 4The relationship diagram ofmultipopulation demonstrat-ing
The final constraint violation of each individual in thepopulation can be obtained by calculating the mean of thenormalized constraint violations
In order to take advantage of the infeasible solutionswith better performance we proposed the following adaptivedifferential mutation operator to generate individual varia-tion learning from the mutation operators in DErand-to-best1bin according to rules defined by Price et al [11]Considering each individual vector 119909119866
119894 a mutant individual
119909119866+1
119894is defined by
119909119866+1
119894
= 119909119866
119894+ 1198651sdot (119909119866
1198911minus 119909119866
1199031)+ 1198652sdot (119909119866
1198912minus 119909119866
1199032) 119877
119862ge rand ( )
119909119866
119894+ 1198651sdot (119910119866
119894minus 119909119866
1199031)+ 1198652sdot (119909119866
1198911minus 119909119866
1199032) 119877
119862lt rand ( )
(9)
where 1199031and 1199032represent different integers and also different
from 119894 randomly generated over [1119873119901] 119865 is the scalingfactor 119909119866
119891119894(119894 = 1 2 119899) is randomly generated from 119875119891
119910119866
119894(119894 = 1 2 119899) is randomly generated from 119875119888 and 119877
119862is
the mutation factor as follows
119877119862= 1198771198880sdot120574119866+ const
120574119866minus1 + const (10)
where 1198771198880is the initial value of the variability factor const is
a small constant to ensure that the fractional is meaningfuland 120574119866 is defined as follows
120574119866=the number of infeasible solutions in 119875119892
119873119901 (11)
32 Framework of the Proposed Algorithm The proposedalgorithm is described as follows
Step 1 (initialization) Generate the population 119875119892 119875119891 and119875119888 of size 119873119901 119873119875
1 and 119873119875
2 Set the value of 119862
119877(crossover
probability)119866max (the number of function evaluations)119866SM(the iterative number of evolution by NM simplex method)119892 = 1 (the current generation number) and positive controlparameter for scaling the difference vectors 119865
1 1198652 Randomly
generate the parent population 119875119892 from the decision spaceSet the 119875119891 and 119875119888 and let the intermediate populations 119878119892and119872119892 be empty
Step 2 (DE reproduction) By (3) and (9) for mutationcrossover and selection an offspring 119878119892 is created Judge theconstraints of all individuals in 119875119892 In accordance with (8)we first calculate constraint violation degree 119881
119894(119883) of all of
the individuals If 119881119894(119883) = 0 the solution is feasible and
preserved to the intermediate set 119872119892 if 119881119894(119883) gt 0 the
solution is infeasible and preserved to the intermediate set 119878119892
Step 3 (simplex method) Apply NM simplex method oper-ator to the present population if 119892 ge 119866max2 Updatethe present population 119872119892 when the number of iterationsexceeds maximum iterations
Step 4 (119875119891 construction) Rank chromosomes in 119872119892 basedon (5) and generate the elitist population 119875119891 (the size is119873119901)from the ranked population119872119892
Step 5 (119875119888 construction) Add the chromosomes in 119878119892 withslight constraint violations to the 119875119888
Step 6 (mixing the population) Combine 119878119892with the existingparent population to form a new set119872119892 Remove the dupli-cate individuals in119872119892 and the existing parent population
Step 7 (evolution) Randomly choose chromosomes from119875119888 119875119892 and 119875119891 Use the adaptive differential mutation anduniform discrete crossover to obtain the offspring population119875119892 + 1
Step 8 (termination) If the stopping criterion ismet stop andoutput the best solution else go to Step 2
33 Experimental Study In this section we choose threeproblems CTP TNK and BNH as shown in Table 3 to testthe proposed method and compare the method with thecurrent CNSGA-II [35]
8 Journal of Applied Mathematics
Table 3 Test functions
Testfunction
Objective functionmin119865(119883) = min[119891
1(119883) 119891
2(119883)]
Constraints Range ofvariable
CTP
1198911(119883) = 119909
1
1198912(119883) = exp(minus
1198911(119883)
119888 (119883))
times41 +
5
sum
119894=2
[119909119894
2minus 10 cos (2120587119909
119894)]
1198921(119883) = cos (120579) (119891
2(119883) minus 119890) minus sin (120579) 119891
1(119883)
1198922(119883) = 119886
1003816100381610038161003816sin 119887120587 [sin (120579) (1198912 (119883) minus 119890)+ cos (120579) 119891
1(119883)]119888
10038161003816100381610038161003816
119889
1198921(119883) ge 119892
2(119883)
0 le 1199091le 1
minus5 le 119909119894le 5
119894 = 2 3 4 5
BNH1198911(119883) = 4119909
1
2+ 41199092
2
1198912(119883) = (119909
1minus 5)2+ (1199092minus 5)2
1198921(119883) = (119909
1minus 5)2+ 1199092
2minus 25
1198922(119883) = minus(119909
1minus 8)2+ (1199092+ 3) + 77
1198921(119883) le 0 119892
2(119883) le 0
0 le 1199091le 5
0 le 1199092le 3
TNK1198911(119883) = 119909
1
1198912(119883) = 119909
2
1198921(119883) = minus119909
1
2minus 1199092
2+ 1
+01 cos(16 arctan(11990911199092
))
1198922(119883) = (119909
1minus 05)
2+ (1199092minus 05)
1198921(119883) le 0 119892
2(119883) minus 05 le 0
0 le 119909119894le 120587
119894 = 1 2
20
15
10
5
00
Feasible solutionInfeasible solution
02 04 06 08 10f1
f2
Pareto-optimal frontier
(a) CTP1
20
15
10
05
0002 04 06 08 10
Feasible solutionInfeasible solution
Pareto-optimal frontier
f1
f2
(b) CTP2
Figure 5 CTP solution space
For CTP problem there are the six parameters 120579 119886 119887 119888119889 and 119890 that must be chosen in a way so that a portion ofthe unconstrained Pareto-optimal region is infeasible Eachconstraint is an implicit non-linear function of decisionvariables Thus it may be difficult to find a number ofsolutions on a non-linear constraint boundary We take twosets of values for six parameters in CTP problem which aredetermined as (1) CTP1 120579 = 01120587 119886 = 40 119887 = 05 119888 = 1119889 = 2 and 119890 = minus2 (2) CTP2 120579 = minus02120587 119886 = 02 119887 = 10119888 = 1 119889 = 6 and 119890 = 1 The Pareto frontiers the feasiblesolution spaces and the infeasible solution spaces are shownin Figure 5
The parameters are initialized as followsThe size of pop-ulation 119875119892 is119873119901 = 200 size of 119875119891 is119873119875
1= 150 size of 119875119888 is
1198731198752= 10 scaling factors 119865
1and 119865
2are randomly generated
within [05 1] cross rate is 119862119877= 06 maximum evolution
generation is 119866max = 300 and number of SM evolutioniterations is 119866SM = 100 All of the proposed algorithmsand CNSGA-II are real coded with equal population size andequal maximum evolution generation Each algorithm runs20 times independently for each test function
Figure 6 shows the result of a random running of theproposed algorithm and NSGA-II the smooth curve ldquomdashrdquorepresents the Pareto frontier and ldquoXrdquo stands for the solutionachieved by the proposed algorithm or NSGA-II
It is obvious that the proposed algorithm returns abetter approximation to the true Pareto-optimal frontie anda distribution of higher uniformity We also evaluated the
Journal of Applied Mathematics 9
40
35
30
25
20
15
10
050
Pareto frontierProposed algorithm
02 0301 04 05 07 0906 08 10
f2
f1
(a) CTP1 calculated by the proposed algorithm
CNSGA-IIPareto frontier
40
35
30
25
20
15
10
05
f2
f1
0 02 0301 04 05 07 0906 08 10
(b) CTP1 calculated by CNSGA-II
1211100908070605040302
f2
Pareto frontierProposed algorithm
f1
0 02 0301 04 05 07 0906 08 10
(c) CTP2 calculated by the proposed algorithm
CNSGA-IIPareto frontier
1211100908070605040302
f2
f1
0 02 0301 04 05 07 0906 08 10
(d) CTP2 calculated by CNSGA-II
50
40
30
20
10
00 50 100 150
Pareto frontierProposed algorithm
f2
f1
(e) BNH calculated by the proposed algorithm
Pareto frontier
50
40
30
20
10
00 50 100 150
f2
f1
CNSGA-II
(f) BNH calculated by CNSGA-II
Figure 6 Continued
10 Journal of Applied Mathematics
15
10
05
00 05 10 15
Pareto frontierProposed algorithm
f2
f1
(g) TNK calculated by the proposed algorithm
Pareto frontier
15
10
05
00 05 10 15
f2
f1
CNSGA-II
(h) TNK calculated by CNSGA-II
Figure 6 The comparison chart of Pareto frontier
Table 4 The comparison of performance
Test function Algorithm 120574 SP
CTP1 CNSGA-Π 0021317 plusmn 0000323 0873321 plusmn 008725
The proposed 0009836 plusmn 0000410 0567933 plusmn 001845
CTP2 CNSGA-Π 0011120 plusmn 0000753 078314 plusmn 002843
The proposed 0007013 plusmn 0000554 0296543 plusmn 000453
BNH CNSGA-Π 0014947 plusmn 0000632 0336941 plusmn 000917
The proposed 0013766 plusmn 0000043 0209321 plusmn 000561
TNK CNSGA-Π 0013235 plusmn 0000740 0464542 plusmn 000730
The proposed 0006435 plusmn 0000017 0224560 plusmn 000159
Feed water Water
Feed ore Ball mill OverflowMixing
PumpClassifierwater
additionFirst
classifierrecycle First-
overflow Spiral classifier 2Second-stage
to flotationSpiral classifier 1 Second classifier recycle(rough concentrate)
addition
pump
stageoverflow
Figure 7 Flow diagram of the grinding and classification process
algorithms based on the two criterions 120574 and SP as shownin Table 4 It can be observed from the data in Table 4that the proposed algorithm performs significantly betterthan the classical CNSGA-II algorithm in convergence anddistribution uniformityThe simulation results show that thisalgorithm can accurately converge to global Pareto solutionsand can maintain diversity of population
4 Optimization of Grinding andClassification Process
41 Bauxite Grinding and Classification Process Thegrindingand classification process is the key preparation for thebauxite mineral processing Here we consider a bauxitegrinding process in a certain mineral company with singlegrinding and two-stage classification as shown in Figure 7
The process consists in a grinding ball mill and twospiral classifiers First classifier recycle will be put back tothe ball mill for regrinding and the first-stage overflowwill be put into second spiral classifier after being mixedwith water the second classifier recycle will be prepared forBayer production as the rough concentrate and the second-stage overflow will be sent to the next flotation processThe production objectives are composed of the productionyields technically represented by the solid flow of feed oresince the process is nonstorable and the mineral processingquality represented by percentage of the small-size fractionsof mineral particles in the second-stage overflows
Journal of Applied Mathematics 11
Table 5 Notations for the model of the grinding and classificationprocess
Notation Description
119901119894
Particle percentage of the 119894th size fraction inthe ball mill overflow
119891119895
Particle percentage of the 119894th size fraction inthe Feed ore
119862 Rate of the first classifier recycle119864119886119894 The efficiency of the first spiral classifier
120591 Themean residence time1198751 The internal concentration in ball mill119872MF The solid flow of feed ore1198821 The water addition of the first classifier recycle
1198822 The classifier water addition
119887119894119895 The breakage distribution function119878119894 The breakage rate function119889119894 The particle with the 119894th size
120572119888119894
The particle percentage of the 119894th size in thefirst classifier overflow
119889119900 The unit size of the particle1198752 First-stage overflow120601119861 Ball filling rate
119860119888 119861119888
Parameters of the first-stage classifier overflowsize fraction distribution
119864119886119894 The efficiency of the first spiral classifier
11988950119888
The particle size fraction after correctionseparation
119898 The separation accuracy119896 The intermix index
1198641198861015840
min 1198891015840
50119888 1198981015840 1198961015840 The corresponding key parameters to theefficiency of the second spiral classifier
119886 120572 120583 ΛFour key parameters to control the breakagerate function
119889min 119889max Theminimum and maximum particle sizes119860MF 119861MF Parameters of feed ore size fraction distribution
119865(119894)The cumulative particle percentage less thanthe 119894th size fraction in feed ore
1205721198881015840
119894
The particle percentage of the 119894th size fractionin the second classifier overflow
1198641198861015840
119894 The efficiency of the second spiral classifier
42 Predictive Model of the Grinding and Classification Pro-cess Here we establish the mathematical predictive modelof each unit process in the bauxite grinding and classificationprocess The notations of the indexes decision variables andparameters are listed in Table 5 These notations will be usedfor the model of the grinding and classification process
421 Ball Mill Circuit Model Here 119901119894is the particle per-
centage of 119894th size fraction in the ball mill overflow 119891119895is the
particle percentage of 119894th size fraction in feed ore rate of thefirst classifier recycle 119862 is known and 119864119886
119894is the efficiency of
the first spiral classifier According to a technical report offield investigation and study we have that
119901119894(1 + 119862) =
119889119894119894119891119894+ sum119894minus1
119895=1119894gt1119889119894119895[119864119886119894119901119895(1 + 119862) + 119891
119895]
1 minus 119889119894119894119864119886119894
119889119894119895=
119890119895 119894 = 119895
119894minus1
sum
119896=119895
119888119894119896119888119895119896(119890119896minus 119890119894) 119894 gt 119895
119888119894119895=
minus
119895minus1
sum
119896=119894
119888119894119896119888119895119896 119894 lt 119895
1 119894 = 119895
1
119878119894minus 119878119895
119894minus1
sum
119896=119895
119878119896119887119894119896119888119896119895 119894 gt 119895
119890119895=
1
(1 + 05 sdot 120591119878119895) (1 + 025 sdot 120591119878
119895)2 120591 =
81198751
119872MF (12)
where 120591 is the mean residence time of minerals 1198751is the
internal concentration in ball mill and
1198751=
119872MF (1 + 119862)
119872MF (1 + 119862) +1198821 + 03 (1198821 +1198822) (13)
where119872MF (119905ℎ) is the solid flow of feed ore1198821is the water
addition of the first classifier recycle and1198822is the classifier
water addition 119887119894119895is the breakage distribution function 119878
119894
is the breakage rate function and it satisfied the followingequation
119878119894=
(119886(119889119894119889119900)120572
)
(1 + (119889119894120583)Λ
)
(14)
where 119889119894is the particle with the 119894th size 119889
119900it is a unit when
per millimeter is a unit 119889119900= 1 119889
119894= 119894 (mm) and 119886 120572 120583
and Λ are four key parameters to control the breakage ratefunction
In a concrete grinding and classification process the ballmill size is fixed and the speed of ball mill is constantThrough data acquisition and testing of grinding and clas-sification steady-state loop the regression model between 119886120572 120583 Λ and condition variables size fraction distributioncan be established The input variables are ball filling rate120601119861 solid flow of feed ore 119872MF water addition of the first
classifier recycle1198821 and parameters of feed ore size fraction
distribution 119860MF 119861MF The regression model is
[119886 120572 120583 Λ]119879
=[[
[
11990911
sdot sdot sdot 1199091119895
d
1199091198941
sdot sdot sdot 119909119894119895
]]
]
sdot [1198821 120601119861 119860MF 119861MF 119872MF]119879
(15)
The value of 119909119894119895(119894 = 1 2 3 4 119895 = 1 2 5) can be
obtained by the experimental data regression 119860MF 119861MF can
12 Journal of Applied Mathematics
be obtained from feed ore size fraction distribution and 119865(119894)is the cumulative particle percentage less than the 119894th sizefraction in feed ore and it is represented as follows
119865 (119894) = 1 minus exp (minus119860MF119889119861MF119894
) (16)
422 Spiral Classifier Model 120572119888119894is the particle percentage
of the 119894th size fraction in the first classifier overflow and 119901119894
is the particle percentage of the 119894th size fraction in ball milloverflow The spiral classifier model is as follows
120572119888119894=
119901119894times (1 minus 119864119886
119894)
sum119894=1
(119901119894times (1 minus 119864119886
119894))times 100 (17)
where 119864119886119894is the efficiency of the first spiral classifier and the
mechanism formula of 119864119886119894is shown as follows
119864119886119894= 1 minus exp [minus0693(
119889119894minus 119889min
11988950119888
minus 119889min)
119898
]
+ 119864119886min sdot [1 minus (119889119894minus 119889min119889max
)]
119896
(18)
where 119889119894is the particle with the 119894th size 119889min and 119889max
represent maximum and minimum particle sizes 11988950119888
isthe particle size fraction after correction separation 119898 isseparation accuracy and 119896 is intermix index
Through data acquisition and testing of grinding andclassification steady-state loop the regressionmodel betweenclassification parameters and condition variables size frac-tion distribution can be established The input variablesinclude the solid flow of feed ore 119872MF the classifier wateraddition 119882
2 and the parameters of ball mill overflow size
fraction distribution 119860MF 119861MF The regression model isshown as follows
[119864119886min 11988950119888
119898 119896]119879
=[[
[
11991011
sdot sdot sdot 1199101119895
d
1199101198941
sdot sdot sdot 119910119894119895
]]
]
sdot [119872MF 1198822119860MP 119861MP]
119879
(19)
where the value of 119910119894119895(119894 = 1 2 3 4 119895 = 1 2 3 4) can be
obtained by data regressionThe first-stage overflow 119875
2calculation formula is as
follows
1198752=
119872MF(119872MF +1198821 +1198822)
(20)
Similarly we can get the second spiral classifier model asfollows
1205721198881015840
119894=
120572119888119894times (1 minus 119864119886
1015840
119894)
sum119894(120572119888119894times (1 minus 1198641198861015840))
times 100 (21)
where 1205721198881015840119894is the particle percentage of the 119894th size fraction in
the second classifier overflow 120572119888119894is the particle percentage
of the 119894th size fraction in the first classifier overflow and
1198641198861015840
119894is the efficiency of the second spiral classifier The spiral
classifier model is as follows
1198641198861015840
119894= 1 minus exp[
[
minus0693(1198891015840
119894minus 1198891015840
min1198891015840
50119888minus 1198891015840
min)
1198981015840
]
]
+ 1198641198861015840
min sdot [1 minus (1198891015840
119894minus 1198891015840
min1198891015840max
)]
1198961015840
(22)
where 1198641198861015840min 1198891015840
50119888 1198981015840 and 119896
1015840 are key parameters to theefficiency of the second spiral classifier Through data acqui-sition and testing of grinding and classification steady-stateloop the regression model between classification parametersand condition variables size fraction distribution can beestablished The input variables include solid flow of feedore119872MF and parameters of the first-stage classifier overflowsize fraction distribution 119860
119888 119861119888 which are solved by similar
equation to (20) The regression model is shown as follows
[1198641198861015840
min 1198891015840
5011988811989810158401198961015840]119879
=
[[[
[
1199101015840
11sdot sdot sdot 1199101015840
1119895
d
1199101015840
1198941sdot sdot sdot 119910
1015840
119894119895
]]]
]
sdot [119872MF 119860119888119861119888]119879
(23)
where the value of1199101015840119894119895(119894 = 1 2 3 4 119895 = 1 2 3) can be obtained
by experimental data regression
43 Optimization Model of Grinding and Classification Pro-cess Two objective functions in the process are identifiedone is tomaximize output119891
1(119883) and the other is tomaximize
the small-size fractions (less than 0075mm fractions) in thesecond-stage overflow 119891
2(119883) It is also necessary to ensure
that the grinding productmeets all of other technical require-ments and the least disturbance in the following flotationcircuit As the constraints the feed load of the grinding circuit119872MF the steel ball filling rate 120601
119861 the first and the second
overflows 1198751and 119875
2 and the particle percentage of fine size
fraction in the first and the second classifier overflows 120572119888minus0075
and 1205721198881015840minus0075
should be within the user specified boundsThe operation variables are the solid flow of feed ore
119872MF water addition of the first classifier recycle 1198821 ball
filling rate 120601119861 and water addition of the second classifier119882
2
Based on all of the above grinding and classification processmultiobjective optimization model is as follows
max 119865 = max [1198911(119883) 119891
2(119883)]
1198911(119883) = 119872MF
1198912(119883) = 120572119888
1015840
minus0075
= 119891 (119872MF11988211198822 120601119861)
st 119872MFmin le 119872MF le 119872MFmax120601119861min le 120601119861 le 120601119861max1198751min le 1198751 le 1198751max1198752min le 1198752 le 1198752max120572119888minus0075
ge 120572119888min
1205721198881015840
minus0075ge 1205721198881015840
min
(24)
Journal of Applied Mathematics 13
Table 6 The optimization results calculated by the proposed algo-rithm
Number 1198911(119883) (th) 119891
2(119883) ()
1 92972 908142 91810 919003 92360 909234 90620 934605 91310 924946 90170 938007 89390 952008 89480 949329 91530 9240010 89298 9576311 89824 9454912 89800 9439513 89710 9461814 91170 9280015 91880 9184016 92190 9128117 89870 9409018 91000 9328519 91960 9152020 91124 93221
85
87
89
91
93
95
97
74000 79000 84000 94000
Optimization resultsThe raw data
f1(x)
f2(x)
()
89000
Figure 8 The comparison chart between optimization results andindustrial data
In (24) 119891(119872MF11988211198822 120601119861) implicates the model ofgrinding and classification represented by (12)ndash(23) Theproposed algorithm is applied to solve the problem and theoptimization results are shown in Table 6
With the practical process data from a grinding circuitof a mineral plant the simulation of this hybrid intelligentmethod adopted the same parameters on the variation infresh slurry feed velocity density particle size distributionand cyclone feed operating configurations
The comparison of production data and optimizationresults in Table 6 is shown in Figure 8 where ldquoXrdquo repre-sents the proposed algorithm optimization results and ldquoIrdquorepresents the original data collected from the field withoutoptimization of raw data According to the objectives thedata point closer to the upper right edge is more beneficialObviously the proposed optimization result is far betterthan the original data indicating the effectiveness of theoptimization approach
44 TOPSIS Method for Solution Selection The resolution ofamultiobjective optimization problemdoes not endwhen thePareto-optimal set is found In practical operational prob-lems a single solution must be selected TOPSIS [36] is auseful technique in dealing with multiattribute or multicrite-ria decision-making (MADMMCDM) problems in the realworld The standard TOPSIS method attempts to choosealternatives that simultaneously have the shortest distancefrom the positive-ideal solution and the farthest distancefrom the negative-ideal solution According to the TOPSISmethod the relative closeness coefficient is calculated andthe best solution in Table 6 is the solution number 10 as1198911(119883) = 89298 (119905ℎ) and 119891
2(119883) = 95763 () The
corresponding decision variables are 119872MF = 89298 (119905ℎ)120601119861= 32119882
1= 75127 (119905ℎ) and119882
2= 16296 (119905ℎ)
5 Conclusions
Promoted by the requirements of engineering optimizationin complex practical processes of grinding and classificationwe proposed a hybrid multiobjective differential evolutionalgorithm with a few beneficial features integrated Firstly anarchiving mechanism for infeasible solutions is establishedwith functional partitioned multi-population which aims todirect the population to approach or land in the feasibleregion from different directions during evolution Secondlywe propose an infeasible constraint violations function toselect infeasible population with better performance sothat they are allowed to be saved and to participate inthe subsequent evolution Thirdly a nondominating rankingstrategy is designed to improve the crowding-distance sortingand return uniform distribution of Pareto solutions Finallythe simplex method is inserted in the differential evolutionprocess to purposefully enrich the diversity without excessivecomputation cost The advantage of the proposed algorithmis the exemption from constructing penalty function andthe preservation of meaningful infeasible solutions directlySimulation results on benchmarks indicate that the proposedalgorithm can converge quickly and effectively to the truePareto frontier with better distribution
Based on the investigated information about grinding cir-cuit process we established a multiobjective optimal modelwith equations frommechanism knowledge parameters rec-ognized by data regression and constraints of technicalrequirements The nonlinear multiobjective optimizationmodel is too complicated to be solved by traditional gradient-based algorithmsThe proposed hybrid differential evolutionalgorithm is applied and tested to achieve a Pareto solution
14 Journal of Applied Mathematics
set It is proven to be valuable for operation decision makingin the industrial process and showed superiority to the oper-ation carried out in the production In fact many operatingparameters in complex processes are highly coupled andconflicting with each other The optimal operation of theentire production process is very difficult to obtain bymanualcalculation let alone the fluctuation situation of processconditions The application case indicates that the proposedmethod has good performance and is helpful to inspirefurther research on evolutionary methods for engineeringoptimization
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgments
This work is supported by the National Natural ScienceFoundation of China (61374156 61273187 and 61134006)the National Key Technology Research and DevelopmentProgram of the Ministry of Science and Technology ofChina (2012BAK09B04) the Science Fund for CreativeResearch Groups of the National Natural Science Foundationof China (61321003) the Fund for Doctor Station of theMinistry of Education (20110162130011 and 20100162120019)and theHunanProvince Science andTechnology PlanProject(2012CK4018)
References
[1] K Mitra and R Gopinath ldquoMultiobjective optimization ofan industrial grinding operation using elitist nondominatedsorting genetic algorithmrdquo Chemical Engineering Science vol59 no 2 pp 385ndash396 2004
[2] G Yu T Y Chai and X C Luo ldquoMultiobjective productionplanning optimization using hybrid evolutionary algorithmsfor mineral processingrdquo IEEE Transactions on EvolutionaryComputation vol 15 no 4 pp 487ndash514 2011
[3] T Y Ma W H Gui Y L Wang and C H Yang ldquoDynamicoptimization control for grinding and classification processrdquoControl andDecision vol 27 no 2 pp 286ndash290 2012 (Chinese)
[4] F Y Edgeworth Mathematical Physics An Essay on the Appli-cation of Mathematics to the Moral Sciences C Kegan Paul ampCompany London UK 1881
[5] V Pareto Cours drsquoEconomie Politique F Rouge LausanneSwitzerland 1896
[6] X F Chen W H Gui Y L Wang and L Cen ldquoMulti-stepoptimal control of complex process a genetic programmingstrategy and its applicationrdquo Engineering Applications of Artifi-cial Intelligence vol 17 no 5 pp 491ndash500 2004
[7] C A C Coello ldquoEvolutionary multi-objective optimizationa historical view of the fieldrdquo IEEE Computational IntelligenceMagazine vol 1 no 1 pp 28ndash36 2006
[8] K Deb A Pratap S Agarwal and T Meyarivan ldquoA fastand elitist multiobjective genetic algorithm NSGA-IIrdquo IEEETransactions on Evolutionary Computation vol 6 no 2 pp 182ndash197 2002
[9] C M Fonesca and P J Fleming ldquoGenetic algorithms formulti objective optimization formulation discussion and gen-eralizationrdquo in Proceedings of the 5th International Conferenceon Genetic Algorithms pp 415ndash423 Morgan Kaufmann SanMateo Calif USA 1993
[10] S Sengupta S Das M Nasir A V Vasilakos and W PedryczldquoEnergy-efficient differentiated coverage of dynamic objectsusing an improved evolutionary multi-objective optimizationalgorithm with fuzzy-dominancerdquo in Proceedings of IEEECongress on Evolutionary Computation pp 1ndash8 IEEE PressBrisbane Australia 2012
[11] K V Price R M Storn and J A Lampinen DifferentialEvolution a Practical Approach to Global Optimization SpringerBerlin Germany 2005
[12] W Y Gong and Z H Cai ldquoAn improved multiobjective dif-ferential evolution based on Pareto-adaptive 120576-dominance andorthogonal designrdquo European Journal of Operational Researchvol 198 no 2 pp 576ndash601 2009
[13] X H Zeng W K Wong and S Y Leung ldquoAn operatorallocation optimization model for balancing control of thehybrid assembly lines using Pareto utility discrete differentialevolution algorithmrdquo Computers and Operations Research vol39 no 5 pp 1145ndash1159 2012
[14] E Zitzler and L Thiele ldquoMultiobjective evolutionary algo-rithms a comparative case study and the strength Paretoapproachrdquo IEEETransactions onEvolutionaryComputation vol3 no 4 pp 257ndash271 1999
[15] A Al-Ani A Alsukker and R N Khushaba ldquoFeature subsetselection using differential evolution and a wheel based searchstrategyrdquo Swarm and Evolutionary Computation vol 9 pp 15ndash26 2013
[16] G Y Li and M G Liu ldquoThe summary of differential evolutionalgorithm and its improvementsrdquo in Proceedings of the 3rdInternational Conference on Advanced Computer Theory andEngineering (ICACTE rsquo10) pp V3153ndashV3156 Chengdu ChinaAugust 2010
[17] H A Abbass ldquoThe self-adaptive Pareto differential evolutionalgorithmrdquo in Proceedings of the IEEE Congress on EvolutionaryComputation pp 831ndash836 IEEE Press Honolulu Hawaii USAMay 2002
[18] N K Madavan ldquoMultiobjective optimization using a Paretodifferential evolution approachrdquo in Proceedings of the IEEECongress on Evolutionary Computation pp 1145ndash1150 IEEEPress Honolulu Hawaii USA May 2002
[19] F Xue A C Sanderson and R J Graves ldquoPareto-based multi-objective differential evolutionrdquo in Proceedings of the IEEECongress on Evolutionary Computation pp 862ndash869 IEEEPress Piscataway NJ USA December 2003
[20] T Robic and B Filipi ldquoDEMO differential evolution formultiobjective optimizationrdquo in Evolutionary Multi-CriterionOptimization pp 520ndash533 Springer Berlin Germany 2005
[21] V L Huang A K Qin P N Suganthan and M F TasgetirenldquoMulti-objective optimization based on self-adaptive differen-tial evolution algorithmrdquo in Proceedings of the IEEE Congresson Evolutionary Computation (CEC rsquo07) pp 3601ndash3608 IEEEPress Singapore September 2007
[22] V L Huang S Z Zhao R Mallipeddi and P N SuganthanldquoMulti-objective optimization using self-adaptive differentialevolution algorithmrdquo in Proceedings of the IEEE Congress onEvolutionary Computation (CEC rsquo09) pp 190ndash194 IEEE PressTrondheim Norway May 2009
Journal of Applied Mathematics 15
[23] J A Adeyemo and F A O Otieno ldquoMulti-objective differentialevolution algorithm for solving engineering problemsrdquo Journalof Applied Sciences vol 9 no 20 pp 3652ndash3661 2009
[24] Y Wang and Z Cai ldquoCombining multiobjective optimizationwith differential evolution to solve constrained optimizationproblemsrdquo IEEE Transactions on Evolutionary Computationvol 16 no 1 pp 117ndash134 2012
[25] A Homaifar C X Qi and S H Lai ldquoConstrained optimizationvia genetic algorithmsrdquo Simulation vol 62 no 4 pp 242ndash2531994
[26] J A Joines and C R Houck ldquoOn the use of non-stationarypenalty functions to solve nonlinear constrained optimizationproblems with GArsquosrdquo in Proceedings of the 1st IEEE Conferenceon Evolutionary Computation pp 579ndash584 Orlando Fla USAJune 1994
[27] F Jimenez and J L Verdegay ldquoEvolutionary techniques forconstrained optimization problemsrdquo in Proceedings of the 7thEuropean Congress on Intelligent Techniques and Soft Computing(EUFIT rsquo99) Aachen Germany 1999
[28] C Sun J Zeng and J Pan ldquoAn improved vector particleswarm optimization for constrained optimization problemsrdquoInformation Sciences vol 181 no 6 pp 1153ndash1163 2011
[29] Z Michalewicz and M Schoenauer ldquoEvolutionary algorithmsfor constrained parameter optimization problemsrdquo Evolution-ary Computation vol 4 no 1 pp 1ndash32 1996
[30] C A C Coello and A D Christiansen ldquoMoses a multiob-jective optimization tool for engineering designrdquo EngineeringOptimization vol 31 no 1ndash3 pp 337ndash368 1999
[31] Y G Woldesenbet G G Yen and B G Tessema ldquoConstrainthandling in multiobjective evolutionary optimizationrdquo IEEETransactions on Evolutionary Computation vol 13 no 3 pp514ndash525 2009
[32] J A Nelder and R Mead ldquoA simplex method for functionminimizationrdquoThe Computer Journal vol 7 no 4 pp 308ndash3131965
[33] D A van Veldhuizen and G B Lamont ldquoMultiobjective evolu-tionary algorithms analyzing the state-of-the-artrdquo EvolutionaryComputation vol 8 no 2 pp 125ndash147 2000
[34] R Farmani D A Savic andG AWalters ldquoEvolutionarymulti-objective optimization in water distribution network designrdquoEngineering Optimization vol 37 no 2 pp 167ndash183 2005
[35] K Deb and T Goel Controlled Elitist Non-Dominated SortingGenetic Algorithms for Better Convergence Springer BerlinGermany 2001
[36] Y J Lai T Y Liu and C L Hwang ldquoTopsis for MODMrdquoEuropean Journal of Operational Research vol 76 no 3 pp486ndash500 1994
Submit your manuscripts athttpwwwhindawicom
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical Problems in Engineering
Hindawi Publishing Corporationhttpwwwhindawicom
Differential EquationsInternational Journal of
Volume 2014
Applied MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Probability and StatisticsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical PhysicsAdvances in
Complex AnalysisJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
OptimizationJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
CombinatoricsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Operations ResearchAdvances in
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Function Spaces
Abstract and Applied AnalysisHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of Mathematics and Mathematical Sciences
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Algebra
Discrete Dynamics in Nature and Society
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Decision SciencesAdvances in
Discrete MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014 Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Stochastic AnalysisInternational Journal of
8 Journal of Applied Mathematics
Table 3 Test functions
Testfunction
Objective functionmin119865(119883) = min[119891
1(119883) 119891
2(119883)]
Constraints Range ofvariable
CTP
1198911(119883) = 119909
1
1198912(119883) = exp(minus
1198911(119883)
119888 (119883))
times41 +
5
sum
119894=2
[119909119894
2minus 10 cos (2120587119909
119894)]
1198921(119883) = cos (120579) (119891
2(119883) minus 119890) minus sin (120579) 119891
1(119883)
1198922(119883) = 119886
1003816100381610038161003816sin 119887120587 [sin (120579) (1198912 (119883) minus 119890)+ cos (120579) 119891
1(119883)]119888
10038161003816100381610038161003816
119889
1198921(119883) ge 119892
2(119883)
0 le 1199091le 1
minus5 le 119909119894le 5
119894 = 2 3 4 5
BNH1198911(119883) = 4119909
1
2+ 41199092
2
1198912(119883) = (119909
1minus 5)2+ (1199092minus 5)2
1198921(119883) = (119909
1minus 5)2+ 1199092
2minus 25
1198922(119883) = minus(119909
1minus 8)2+ (1199092+ 3) + 77
1198921(119883) le 0 119892
2(119883) le 0
0 le 1199091le 5
0 le 1199092le 3
TNK1198911(119883) = 119909
1
1198912(119883) = 119909
2
1198921(119883) = minus119909
1
2minus 1199092
2+ 1
+01 cos(16 arctan(11990911199092
))
1198922(119883) = (119909
1minus 05)
2+ (1199092minus 05)
1198921(119883) le 0 119892
2(119883) minus 05 le 0
0 le 119909119894le 120587
119894 = 1 2
20
15
10
5
00
Feasible solutionInfeasible solution
02 04 06 08 10f1
f2
Pareto-optimal frontier
(a) CTP1
20
15
10
05
0002 04 06 08 10
Feasible solutionInfeasible solution
Pareto-optimal frontier
f1
f2
(b) CTP2
Figure 5 CTP solution space
For CTP problem there are the six parameters 120579 119886 119887 119888119889 and 119890 that must be chosen in a way so that a portion ofthe unconstrained Pareto-optimal region is infeasible Eachconstraint is an implicit non-linear function of decisionvariables Thus it may be difficult to find a number ofsolutions on a non-linear constraint boundary We take twosets of values for six parameters in CTP problem which aredetermined as (1) CTP1 120579 = 01120587 119886 = 40 119887 = 05 119888 = 1119889 = 2 and 119890 = minus2 (2) CTP2 120579 = minus02120587 119886 = 02 119887 = 10119888 = 1 119889 = 6 and 119890 = 1 The Pareto frontiers the feasiblesolution spaces and the infeasible solution spaces are shownin Figure 5
The parameters are initialized as followsThe size of pop-ulation 119875119892 is119873119901 = 200 size of 119875119891 is119873119875
1= 150 size of 119875119888 is
1198731198752= 10 scaling factors 119865
1and 119865
2are randomly generated
within [05 1] cross rate is 119862119877= 06 maximum evolution
generation is 119866max = 300 and number of SM evolutioniterations is 119866SM = 100 All of the proposed algorithmsand CNSGA-II are real coded with equal population size andequal maximum evolution generation Each algorithm runs20 times independently for each test function
Figure 6 shows the result of a random running of theproposed algorithm and NSGA-II the smooth curve ldquomdashrdquorepresents the Pareto frontier and ldquoXrdquo stands for the solutionachieved by the proposed algorithm or NSGA-II
It is obvious that the proposed algorithm returns abetter approximation to the true Pareto-optimal frontie anda distribution of higher uniformity We also evaluated the
Journal of Applied Mathematics 9
40
35
30
25
20
15
10
050
Pareto frontierProposed algorithm
02 0301 04 05 07 0906 08 10
f2
f1
(a) CTP1 calculated by the proposed algorithm
CNSGA-IIPareto frontier
40
35
30
25
20
15
10
05
f2
f1
0 02 0301 04 05 07 0906 08 10
(b) CTP1 calculated by CNSGA-II
1211100908070605040302
f2
Pareto frontierProposed algorithm
f1
0 02 0301 04 05 07 0906 08 10
(c) CTP2 calculated by the proposed algorithm
CNSGA-IIPareto frontier
1211100908070605040302
f2
f1
0 02 0301 04 05 07 0906 08 10
(d) CTP2 calculated by CNSGA-II
50
40
30
20
10
00 50 100 150
Pareto frontierProposed algorithm
f2
f1
(e) BNH calculated by the proposed algorithm
Pareto frontier
50
40
30
20
10
00 50 100 150
f2
f1
CNSGA-II
(f) BNH calculated by CNSGA-II
Figure 6 Continued
10 Journal of Applied Mathematics
15
10
05
00 05 10 15
Pareto frontierProposed algorithm
f2
f1
(g) TNK calculated by the proposed algorithm
Pareto frontier
15
10
05
00 05 10 15
f2
f1
CNSGA-II
(h) TNK calculated by CNSGA-II
Figure 6 The comparison chart of Pareto frontier
Table 4 The comparison of performance
Test function Algorithm 120574 SP
CTP1 CNSGA-Π 0021317 plusmn 0000323 0873321 plusmn 008725
The proposed 0009836 plusmn 0000410 0567933 plusmn 001845
CTP2 CNSGA-Π 0011120 plusmn 0000753 078314 plusmn 002843
The proposed 0007013 plusmn 0000554 0296543 plusmn 000453
BNH CNSGA-Π 0014947 plusmn 0000632 0336941 plusmn 000917
The proposed 0013766 plusmn 0000043 0209321 plusmn 000561
TNK CNSGA-Π 0013235 plusmn 0000740 0464542 plusmn 000730
The proposed 0006435 plusmn 0000017 0224560 plusmn 000159
Feed water Water
Feed ore Ball mill OverflowMixing
PumpClassifierwater
additionFirst
classifierrecycle First-
overflow Spiral classifier 2Second-stage
to flotationSpiral classifier 1 Second classifier recycle(rough concentrate)
addition
pump
stageoverflow
Figure 7 Flow diagram of the grinding and classification process
algorithms based on the two criterions 120574 and SP as shownin Table 4 It can be observed from the data in Table 4that the proposed algorithm performs significantly betterthan the classical CNSGA-II algorithm in convergence anddistribution uniformityThe simulation results show that thisalgorithm can accurately converge to global Pareto solutionsand can maintain diversity of population
4 Optimization of Grinding andClassification Process
41 Bauxite Grinding and Classification Process Thegrindingand classification process is the key preparation for thebauxite mineral processing Here we consider a bauxitegrinding process in a certain mineral company with singlegrinding and two-stage classification as shown in Figure 7
The process consists in a grinding ball mill and twospiral classifiers First classifier recycle will be put back tothe ball mill for regrinding and the first-stage overflowwill be put into second spiral classifier after being mixedwith water the second classifier recycle will be prepared forBayer production as the rough concentrate and the second-stage overflow will be sent to the next flotation processThe production objectives are composed of the productionyields technically represented by the solid flow of feed oresince the process is nonstorable and the mineral processingquality represented by percentage of the small-size fractionsof mineral particles in the second-stage overflows
Journal of Applied Mathematics 11
Table 5 Notations for the model of the grinding and classificationprocess
Notation Description
119901119894
Particle percentage of the 119894th size fraction inthe ball mill overflow
119891119895
Particle percentage of the 119894th size fraction inthe Feed ore
119862 Rate of the first classifier recycle119864119886119894 The efficiency of the first spiral classifier
120591 Themean residence time1198751 The internal concentration in ball mill119872MF The solid flow of feed ore1198821 The water addition of the first classifier recycle
1198822 The classifier water addition
119887119894119895 The breakage distribution function119878119894 The breakage rate function119889119894 The particle with the 119894th size
120572119888119894
The particle percentage of the 119894th size in thefirst classifier overflow
119889119900 The unit size of the particle1198752 First-stage overflow120601119861 Ball filling rate
119860119888 119861119888
Parameters of the first-stage classifier overflowsize fraction distribution
119864119886119894 The efficiency of the first spiral classifier
11988950119888
The particle size fraction after correctionseparation
119898 The separation accuracy119896 The intermix index
1198641198861015840
min 1198891015840
50119888 1198981015840 1198961015840 The corresponding key parameters to theefficiency of the second spiral classifier
119886 120572 120583 ΛFour key parameters to control the breakagerate function
119889min 119889max Theminimum and maximum particle sizes119860MF 119861MF Parameters of feed ore size fraction distribution
119865(119894)The cumulative particle percentage less thanthe 119894th size fraction in feed ore
1205721198881015840
119894
The particle percentage of the 119894th size fractionin the second classifier overflow
1198641198861015840
119894 The efficiency of the second spiral classifier
42 Predictive Model of the Grinding and Classification Pro-cess Here we establish the mathematical predictive modelof each unit process in the bauxite grinding and classificationprocess The notations of the indexes decision variables andparameters are listed in Table 5 These notations will be usedfor the model of the grinding and classification process
421 Ball Mill Circuit Model Here 119901119894is the particle per-
centage of 119894th size fraction in the ball mill overflow 119891119895is the
particle percentage of 119894th size fraction in feed ore rate of thefirst classifier recycle 119862 is known and 119864119886
119894is the efficiency of
the first spiral classifier According to a technical report offield investigation and study we have that
119901119894(1 + 119862) =
119889119894119894119891119894+ sum119894minus1
119895=1119894gt1119889119894119895[119864119886119894119901119895(1 + 119862) + 119891
119895]
1 minus 119889119894119894119864119886119894
119889119894119895=
119890119895 119894 = 119895
119894minus1
sum
119896=119895
119888119894119896119888119895119896(119890119896minus 119890119894) 119894 gt 119895
119888119894119895=
minus
119895minus1
sum
119896=119894
119888119894119896119888119895119896 119894 lt 119895
1 119894 = 119895
1
119878119894minus 119878119895
119894minus1
sum
119896=119895
119878119896119887119894119896119888119896119895 119894 gt 119895
119890119895=
1
(1 + 05 sdot 120591119878119895) (1 + 025 sdot 120591119878
119895)2 120591 =
81198751
119872MF (12)
where 120591 is the mean residence time of minerals 1198751is the
internal concentration in ball mill and
1198751=
119872MF (1 + 119862)
119872MF (1 + 119862) +1198821 + 03 (1198821 +1198822) (13)
where119872MF (119905ℎ) is the solid flow of feed ore1198821is the water
addition of the first classifier recycle and1198822is the classifier
water addition 119887119894119895is the breakage distribution function 119878
119894
is the breakage rate function and it satisfied the followingequation
119878119894=
(119886(119889119894119889119900)120572
)
(1 + (119889119894120583)Λ
)
(14)
where 119889119894is the particle with the 119894th size 119889
119900it is a unit when
per millimeter is a unit 119889119900= 1 119889
119894= 119894 (mm) and 119886 120572 120583
and Λ are four key parameters to control the breakage ratefunction
In a concrete grinding and classification process the ballmill size is fixed and the speed of ball mill is constantThrough data acquisition and testing of grinding and clas-sification steady-state loop the regression model between 119886120572 120583 Λ and condition variables size fraction distributioncan be established The input variables are ball filling rate120601119861 solid flow of feed ore 119872MF water addition of the first
classifier recycle1198821 and parameters of feed ore size fraction
distribution 119860MF 119861MF The regression model is
[119886 120572 120583 Λ]119879
=[[
[
11990911
sdot sdot sdot 1199091119895
d
1199091198941
sdot sdot sdot 119909119894119895
]]
]
sdot [1198821 120601119861 119860MF 119861MF 119872MF]119879
(15)
The value of 119909119894119895(119894 = 1 2 3 4 119895 = 1 2 5) can be
obtained by the experimental data regression 119860MF 119861MF can
12 Journal of Applied Mathematics
be obtained from feed ore size fraction distribution and 119865(119894)is the cumulative particle percentage less than the 119894th sizefraction in feed ore and it is represented as follows
119865 (119894) = 1 minus exp (minus119860MF119889119861MF119894
) (16)
422 Spiral Classifier Model 120572119888119894is the particle percentage
of the 119894th size fraction in the first classifier overflow and 119901119894
is the particle percentage of the 119894th size fraction in ball milloverflow The spiral classifier model is as follows
120572119888119894=
119901119894times (1 minus 119864119886
119894)
sum119894=1
(119901119894times (1 minus 119864119886
119894))times 100 (17)
where 119864119886119894is the efficiency of the first spiral classifier and the
mechanism formula of 119864119886119894is shown as follows
119864119886119894= 1 minus exp [minus0693(
119889119894minus 119889min
11988950119888
minus 119889min)
119898
]
+ 119864119886min sdot [1 minus (119889119894minus 119889min119889max
)]
119896
(18)
where 119889119894is the particle with the 119894th size 119889min and 119889max
represent maximum and minimum particle sizes 11988950119888
isthe particle size fraction after correction separation 119898 isseparation accuracy and 119896 is intermix index
Through data acquisition and testing of grinding andclassification steady-state loop the regressionmodel betweenclassification parameters and condition variables size frac-tion distribution can be established The input variablesinclude the solid flow of feed ore 119872MF the classifier wateraddition 119882
2 and the parameters of ball mill overflow size
fraction distribution 119860MF 119861MF The regression model isshown as follows
[119864119886min 11988950119888
119898 119896]119879
=[[
[
11991011
sdot sdot sdot 1199101119895
d
1199101198941
sdot sdot sdot 119910119894119895
]]
]
sdot [119872MF 1198822119860MP 119861MP]
119879
(19)
where the value of 119910119894119895(119894 = 1 2 3 4 119895 = 1 2 3 4) can be
obtained by data regressionThe first-stage overflow 119875
2calculation formula is as
follows
1198752=
119872MF(119872MF +1198821 +1198822)
(20)
Similarly we can get the second spiral classifier model asfollows
1205721198881015840
119894=
120572119888119894times (1 minus 119864119886
1015840
119894)
sum119894(120572119888119894times (1 minus 1198641198861015840))
times 100 (21)
where 1205721198881015840119894is the particle percentage of the 119894th size fraction in
the second classifier overflow 120572119888119894is the particle percentage
of the 119894th size fraction in the first classifier overflow and
1198641198861015840
119894is the efficiency of the second spiral classifier The spiral
classifier model is as follows
1198641198861015840
119894= 1 minus exp[
[
minus0693(1198891015840
119894minus 1198891015840
min1198891015840
50119888minus 1198891015840
min)
1198981015840
]
]
+ 1198641198861015840
min sdot [1 minus (1198891015840
119894minus 1198891015840
min1198891015840max
)]
1198961015840
(22)
where 1198641198861015840min 1198891015840
50119888 1198981015840 and 119896
1015840 are key parameters to theefficiency of the second spiral classifier Through data acqui-sition and testing of grinding and classification steady-stateloop the regression model between classification parametersand condition variables size fraction distribution can beestablished The input variables include solid flow of feedore119872MF and parameters of the first-stage classifier overflowsize fraction distribution 119860
119888 119861119888 which are solved by similar
equation to (20) The regression model is shown as follows
[1198641198861015840
min 1198891015840
5011988811989810158401198961015840]119879
=
[[[
[
1199101015840
11sdot sdot sdot 1199101015840
1119895
d
1199101015840
1198941sdot sdot sdot 119910
1015840
119894119895
]]]
]
sdot [119872MF 119860119888119861119888]119879
(23)
where the value of1199101015840119894119895(119894 = 1 2 3 4 119895 = 1 2 3) can be obtained
by experimental data regression
43 Optimization Model of Grinding and Classification Pro-cess Two objective functions in the process are identifiedone is tomaximize output119891
1(119883) and the other is tomaximize
the small-size fractions (less than 0075mm fractions) in thesecond-stage overflow 119891
2(119883) It is also necessary to ensure
that the grinding productmeets all of other technical require-ments and the least disturbance in the following flotationcircuit As the constraints the feed load of the grinding circuit119872MF the steel ball filling rate 120601
119861 the first and the second
overflows 1198751and 119875
2 and the particle percentage of fine size
fraction in the first and the second classifier overflows 120572119888minus0075
and 1205721198881015840minus0075
should be within the user specified boundsThe operation variables are the solid flow of feed ore
119872MF water addition of the first classifier recycle 1198821 ball
filling rate 120601119861 and water addition of the second classifier119882
2
Based on all of the above grinding and classification processmultiobjective optimization model is as follows
max 119865 = max [1198911(119883) 119891
2(119883)]
1198911(119883) = 119872MF
1198912(119883) = 120572119888
1015840
minus0075
= 119891 (119872MF11988211198822 120601119861)
st 119872MFmin le 119872MF le 119872MFmax120601119861min le 120601119861 le 120601119861max1198751min le 1198751 le 1198751max1198752min le 1198752 le 1198752max120572119888minus0075
ge 120572119888min
1205721198881015840
minus0075ge 1205721198881015840
min
(24)
Journal of Applied Mathematics 13
Table 6 The optimization results calculated by the proposed algo-rithm
Number 1198911(119883) (th) 119891
2(119883) ()
1 92972 908142 91810 919003 92360 909234 90620 934605 91310 924946 90170 938007 89390 952008 89480 949329 91530 9240010 89298 9576311 89824 9454912 89800 9439513 89710 9461814 91170 9280015 91880 9184016 92190 9128117 89870 9409018 91000 9328519 91960 9152020 91124 93221
85
87
89
91
93
95
97
74000 79000 84000 94000
Optimization resultsThe raw data
f1(x)
f2(x)
()
89000
Figure 8 The comparison chart between optimization results andindustrial data
In (24) 119891(119872MF11988211198822 120601119861) implicates the model ofgrinding and classification represented by (12)ndash(23) Theproposed algorithm is applied to solve the problem and theoptimization results are shown in Table 6
With the practical process data from a grinding circuitof a mineral plant the simulation of this hybrid intelligentmethod adopted the same parameters on the variation infresh slurry feed velocity density particle size distributionand cyclone feed operating configurations
The comparison of production data and optimizationresults in Table 6 is shown in Figure 8 where ldquoXrdquo repre-sents the proposed algorithm optimization results and ldquoIrdquorepresents the original data collected from the field withoutoptimization of raw data According to the objectives thedata point closer to the upper right edge is more beneficialObviously the proposed optimization result is far betterthan the original data indicating the effectiveness of theoptimization approach
44 TOPSIS Method for Solution Selection The resolution ofamultiobjective optimization problemdoes not endwhen thePareto-optimal set is found In practical operational prob-lems a single solution must be selected TOPSIS [36] is auseful technique in dealing with multiattribute or multicrite-ria decision-making (MADMMCDM) problems in the realworld The standard TOPSIS method attempts to choosealternatives that simultaneously have the shortest distancefrom the positive-ideal solution and the farthest distancefrom the negative-ideal solution According to the TOPSISmethod the relative closeness coefficient is calculated andthe best solution in Table 6 is the solution number 10 as1198911(119883) = 89298 (119905ℎ) and 119891
2(119883) = 95763 () The
corresponding decision variables are 119872MF = 89298 (119905ℎ)120601119861= 32119882
1= 75127 (119905ℎ) and119882
2= 16296 (119905ℎ)
5 Conclusions
Promoted by the requirements of engineering optimizationin complex practical processes of grinding and classificationwe proposed a hybrid multiobjective differential evolutionalgorithm with a few beneficial features integrated Firstly anarchiving mechanism for infeasible solutions is establishedwith functional partitioned multi-population which aims todirect the population to approach or land in the feasibleregion from different directions during evolution Secondlywe propose an infeasible constraint violations function toselect infeasible population with better performance sothat they are allowed to be saved and to participate inthe subsequent evolution Thirdly a nondominating rankingstrategy is designed to improve the crowding-distance sortingand return uniform distribution of Pareto solutions Finallythe simplex method is inserted in the differential evolutionprocess to purposefully enrich the diversity without excessivecomputation cost The advantage of the proposed algorithmis the exemption from constructing penalty function andthe preservation of meaningful infeasible solutions directlySimulation results on benchmarks indicate that the proposedalgorithm can converge quickly and effectively to the truePareto frontier with better distribution
Based on the investigated information about grinding cir-cuit process we established a multiobjective optimal modelwith equations frommechanism knowledge parameters rec-ognized by data regression and constraints of technicalrequirements The nonlinear multiobjective optimizationmodel is too complicated to be solved by traditional gradient-based algorithmsThe proposed hybrid differential evolutionalgorithm is applied and tested to achieve a Pareto solution
14 Journal of Applied Mathematics
set It is proven to be valuable for operation decision makingin the industrial process and showed superiority to the oper-ation carried out in the production In fact many operatingparameters in complex processes are highly coupled andconflicting with each other The optimal operation of theentire production process is very difficult to obtain bymanualcalculation let alone the fluctuation situation of processconditions The application case indicates that the proposedmethod has good performance and is helpful to inspirefurther research on evolutionary methods for engineeringoptimization
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgments
This work is supported by the National Natural ScienceFoundation of China (61374156 61273187 and 61134006)the National Key Technology Research and DevelopmentProgram of the Ministry of Science and Technology ofChina (2012BAK09B04) the Science Fund for CreativeResearch Groups of the National Natural Science Foundationof China (61321003) the Fund for Doctor Station of theMinistry of Education (20110162130011 and 20100162120019)and theHunanProvince Science andTechnology PlanProject(2012CK4018)
References
[1] K Mitra and R Gopinath ldquoMultiobjective optimization ofan industrial grinding operation using elitist nondominatedsorting genetic algorithmrdquo Chemical Engineering Science vol59 no 2 pp 385ndash396 2004
[2] G Yu T Y Chai and X C Luo ldquoMultiobjective productionplanning optimization using hybrid evolutionary algorithmsfor mineral processingrdquo IEEE Transactions on EvolutionaryComputation vol 15 no 4 pp 487ndash514 2011
[3] T Y Ma W H Gui Y L Wang and C H Yang ldquoDynamicoptimization control for grinding and classification processrdquoControl andDecision vol 27 no 2 pp 286ndash290 2012 (Chinese)
[4] F Y Edgeworth Mathematical Physics An Essay on the Appli-cation of Mathematics to the Moral Sciences C Kegan Paul ampCompany London UK 1881
[5] V Pareto Cours drsquoEconomie Politique F Rouge LausanneSwitzerland 1896
[6] X F Chen W H Gui Y L Wang and L Cen ldquoMulti-stepoptimal control of complex process a genetic programmingstrategy and its applicationrdquo Engineering Applications of Artifi-cial Intelligence vol 17 no 5 pp 491ndash500 2004
[7] C A C Coello ldquoEvolutionary multi-objective optimizationa historical view of the fieldrdquo IEEE Computational IntelligenceMagazine vol 1 no 1 pp 28ndash36 2006
[8] K Deb A Pratap S Agarwal and T Meyarivan ldquoA fastand elitist multiobjective genetic algorithm NSGA-IIrdquo IEEETransactions on Evolutionary Computation vol 6 no 2 pp 182ndash197 2002
[9] C M Fonesca and P J Fleming ldquoGenetic algorithms formulti objective optimization formulation discussion and gen-eralizationrdquo in Proceedings of the 5th International Conferenceon Genetic Algorithms pp 415ndash423 Morgan Kaufmann SanMateo Calif USA 1993
[10] S Sengupta S Das M Nasir A V Vasilakos and W PedryczldquoEnergy-efficient differentiated coverage of dynamic objectsusing an improved evolutionary multi-objective optimizationalgorithm with fuzzy-dominancerdquo in Proceedings of IEEECongress on Evolutionary Computation pp 1ndash8 IEEE PressBrisbane Australia 2012
[11] K V Price R M Storn and J A Lampinen DifferentialEvolution a Practical Approach to Global Optimization SpringerBerlin Germany 2005
[12] W Y Gong and Z H Cai ldquoAn improved multiobjective dif-ferential evolution based on Pareto-adaptive 120576-dominance andorthogonal designrdquo European Journal of Operational Researchvol 198 no 2 pp 576ndash601 2009
[13] X H Zeng W K Wong and S Y Leung ldquoAn operatorallocation optimization model for balancing control of thehybrid assembly lines using Pareto utility discrete differentialevolution algorithmrdquo Computers and Operations Research vol39 no 5 pp 1145ndash1159 2012
[14] E Zitzler and L Thiele ldquoMultiobjective evolutionary algo-rithms a comparative case study and the strength Paretoapproachrdquo IEEETransactions onEvolutionaryComputation vol3 no 4 pp 257ndash271 1999
[15] A Al-Ani A Alsukker and R N Khushaba ldquoFeature subsetselection using differential evolution and a wheel based searchstrategyrdquo Swarm and Evolutionary Computation vol 9 pp 15ndash26 2013
[16] G Y Li and M G Liu ldquoThe summary of differential evolutionalgorithm and its improvementsrdquo in Proceedings of the 3rdInternational Conference on Advanced Computer Theory andEngineering (ICACTE rsquo10) pp V3153ndashV3156 Chengdu ChinaAugust 2010
[17] H A Abbass ldquoThe self-adaptive Pareto differential evolutionalgorithmrdquo in Proceedings of the IEEE Congress on EvolutionaryComputation pp 831ndash836 IEEE Press Honolulu Hawaii USAMay 2002
[18] N K Madavan ldquoMultiobjective optimization using a Paretodifferential evolution approachrdquo in Proceedings of the IEEECongress on Evolutionary Computation pp 1145ndash1150 IEEEPress Honolulu Hawaii USA May 2002
[19] F Xue A C Sanderson and R J Graves ldquoPareto-based multi-objective differential evolutionrdquo in Proceedings of the IEEECongress on Evolutionary Computation pp 862ndash869 IEEEPress Piscataway NJ USA December 2003
[20] T Robic and B Filipi ldquoDEMO differential evolution formultiobjective optimizationrdquo in Evolutionary Multi-CriterionOptimization pp 520ndash533 Springer Berlin Germany 2005
[21] V L Huang A K Qin P N Suganthan and M F TasgetirenldquoMulti-objective optimization based on self-adaptive differen-tial evolution algorithmrdquo in Proceedings of the IEEE Congresson Evolutionary Computation (CEC rsquo07) pp 3601ndash3608 IEEEPress Singapore September 2007
[22] V L Huang S Z Zhao R Mallipeddi and P N SuganthanldquoMulti-objective optimization using self-adaptive differentialevolution algorithmrdquo in Proceedings of the IEEE Congress onEvolutionary Computation (CEC rsquo09) pp 190ndash194 IEEE PressTrondheim Norway May 2009
Journal of Applied Mathematics 15
[23] J A Adeyemo and F A O Otieno ldquoMulti-objective differentialevolution algorithm for solving engineering problemsrdquo Journalof Applied Sciences vol 9 no 20 pp 3652ndash3661 2009
[24] Y Wang and Z Cai ldquoCombining multiobjective optimizationwith differential evolution to solve constrained optimizationproblemsrdquo IEEE Transactions on Evolutionary Computationvol 16 no 1 pp 117ndash134 2012
[25] A Homaifar C X Qi and S H Lai ldquoConstrained optimizationvia genetic algorithmsrdquo Simulation vol 62 no 4 pp 242ndash2531994
[26] J A Joines and C R Houck ldquoOn the use of non-stationarypenalty functions to solve nonlinear constrained optimizationproblems with GArsquosrdquo in Proceedings of the 1st IEEE Conferenceon Evolutionary Computation pp 579ndash584 Orlando Fla USAJune 1994
[27] F Jimenez and J L Verdegay ldquoEvolutionary techniques forconstrained optimization problemsrdquo in Proceedings of the 7thEuropean Congress on Intelligent Techniques and Soft Computing(EUFIT rsquo99) Aachen Germany 1999
[28] C Sun J Zeng and J Pan ldquoAn improved vector particleswarm optimization for constrained optimization problemsrdquoInformation Sciences vol 181 no 6 pp 1153ndash1163 2011
[29] Z Michalewicz and M Schoenauer ldquoEvolutionary algorithmsfor constrained parameter optimization problemsrdquo Evolution-ary Computation vol 4 no 1 pp 1ndash32 1996
[30] C A C Coello and A D Christiansen ldquoMoses a multiob-jective optimization tool for engineering designrdquo EngineeringOptimization vol 31 no 1ndash3 pp 337ndash368 1999
[31] Y G Woldesenbet G G Yen and B G Tessema ldquoConstrainthandling in multiobjective evolutionary optimizationrdquo IEEETransactions on Evolutionary Computation vol 13 no 3 pp514ndash525 2009
[32] J A Nelder and R Mead ldquoA simplex method for functionminimizationrdquoThe Computer Journal vol 7 no 4 pp 308ndash3131965
[33] D A van Veldhuizen and G B Lamont ldquoMultiobjective evolu-tionary algorithms analyzing the state-of-the-artrdquo EvolutionaryComputation vol 8 no 2 pp 125ndash147 2000
[34] R Farmani D A Savic andG AWalters ldquoEvolutionarymulti-objective optimization in water distribution network designrdquoEngineering Optimization vol 37 no 2 pp 167ndash183 2005
[35] K Deb and T Goel Controlled Elitist Non-Dominated SortingGenetic Algorithms for Better Convergence Springer BerlinGermany 2001
[36] Y J Lai T Y Liu and C L Hwang ldquoTopsis for MODMrdquoEuropean Journal of Operational Research vol 76 no 3 pp486ndash500 1994
Submit your manuscripts athttpwwwhindawicom
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical Problems in Engineering
Hindawi Publishing Corporationhttpwwwhindawicom
Differential EquationsInternational Journal of
Volume 2014
Applied MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Probability and StatisticsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical PhysicsAdvances in
Complex AnalysisJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
OptimizationJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
CombinatoricsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Operations ResearchAdvances in
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Function Spaces
Abstract and Applied AnalysisHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of Mathematics and Mathematical Sciences
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Algebra
Discrete Dynamics in Nature and Society
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Decision SciencesAdvances in
Discrete MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014 Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Stochastic AnalysisInternational Journal of
Journal of Applied Mathematics 9
40
35
30
25
20
15
10
050
Pareto frontierProposed algorithm
02 0301 04 05 07 0906 08 10
f2
f1
(a) CTP1 calculated by the proposed algorithm
CNSGA-IIPareto frontier
40
35
30
25
20
15
10
05
f2
f1
0 02 0301 04 05 07 0906 08 10
(b) CTP1 calculated by CNSGA-II
1211100908070605040302
f2
Pareto frontierProposed algorithm
f1
0 02 0301 04 05 07 0906 08 10
(c) CTP2 calculated by the proposed algorithm
CNSGA-IIPareto frontier
1211100908070605040302
f2
f1
0 02 0301 04 05 07 0906 08 10
(d) CTP2 calculated by CNSGA-II
50
40
30
20
10
00 50 100 150
Pareto frontierProposed algorithm
f2
f1
(e) BNH calculated by the proposed algorithm
Pareto frontier
50
40
30
20
10
00 50 100 150
f2
f1
CNSGA-II
(f) BNH calculated by CNSGA-II
Figure 6 Continued
10 Journal of Applied Mathematics
15
10
05
00 05 10 15
Pareto frontierProposed algorithm
f2
f1
(g) TNK calculated by the proposed algorithm
Pareto frontier
15
10
05
00 05 10 15
f2
f1
CNSGA-II
(h) TNK calculated by CNSGA-II
Figure 6 The comparison chart of Pareto frontier
Table 4 The comparison of performance
Test function Algorithm 120574 SP
CTP1 CNSGA-Π 0021317 plusmn 0000323 0873321 plusmn 008725
The proposed 0009836 plusmn 0000410 0567933 plusmn 001845
CTP2 CNSGA-Π 0011120 plusmn 0000753 078314 plusmn 002843
The proposed 0007013 plusmn 0000554 0296543 plusmn 000453
BNH CNSGA-Π 0014947 plusmn 0000632 0336941 plusmn 000917
The proposed 0013766 plusmn 0000043 0209321 plusmn 000561
TNK CNSGA-Π 0013235 plusmn 0000740 0464542 plusmn 000730
The proposed 0006435 plusmn 0000017 0224560 plusmn 000159
Feed water Water
Feed ore Ball mill OverflowMixing
PumpClassifierwater
additionFirst
classifierrecycle First-
overflow Spiral classifier 2Second-stage
to flotationSpiral classifier 1 Second classifier recycle(rough concentrate)
addition
pump
stageoverflow
Figure 7 Flow diagram of the grinding and classification process
algorithms based on the two criterions 120574 and SP as shownin Table 4 It can be observed from the data in Table 4that the proposed algorithm performs significantly betterthan the classical CNSGA-II algorithm in convergence anddistribution uniformityThe simulation results show that thisalgorithm can accurately converge to global Pareto solutionsand can maintain diversity of population
4 Optimization of Grinding andClassification Process
41 Bauxite Grinding and Classification Process Thegrindingand classification process is the key preparation for thebauxite mineral processing Here we consider a bauxitegrinding process in a certain mineral company with singlegrinding and two-stage classification as shown in Figure 7
The process consists in a grinding ball mill and twospiral classifiers First classifier recycle will be put back tothe ball mill for regrinding and the first-stage overflowwill be put into second spiral classifier after being mixedwith water the second classifier recycle will be prepared forBayer production as the rough concentrate and the second-stage overflow will be sent to the next flotation processThe production objectives are composed of the productionyields technically represented by the solid flow of feed oresince the process is nonstorable and the mineral processingquality represented by percentage of the small-size fractionsof mineral particles in the second-stage overflows
Journal of Applied Mathematics 11
Table 5 Notations for the model of the grinding and classificationprocess
Notation Description
119901119894
Particle percentage of the 119894th size fraction inthe ball mill overflow
119891119895
Particle percentage of the 119894th size fraction inthe Feed ore
119862 Rate of the first classifier recycle119864119886119894 The efficiency of the first spiral classifier
120591 Themean residence time1198751 The internal concentration in ball mill119872MF The solid flow of feed ore1198821 The water addition of the first classifier recycle
1198822 The classifier water addition
119887119894119895 The breakage distribution function119878119894 The breakage rate function119889119894 The particle with the 119894th size
120572119888119894
The particle percentage of the 119894th size in thefirst classifier overflow
119889119900 The unit size of the particle1198752 First-stage overflow120601119861 Ball filling rate
119860119888 119861119888
Parameters of the first-stage classifier overflowsize fraction distribution
119864119886119894 The efficiency of the first spiral classifier
11988950119888
The particle size fraction after correctionseparation
119898 The separation accuracy119896 The intermix index
1198641198861015840
min 1198891015840
50119888 1198981015840 1198961015840 The corresponding key parameters to theefficiency of the second spiral classifier
119886 120572 120583 ΛFour key parameters to control the breakagerate function
119889min 119889max Theminimum and maximum particle sizes119860MF 119861MF Parameters of feed ore size fraction distribution
119865(119894)The cumulative particle percentage less thanthe 119894th size fraction in feed ore
1205721198881015840
119894
The particle percentage of the 119894th size fractionin the second classifier overflow
1198641198861015840
119894 The efficiency of the second spiral classifier
42 Predictive Model of the Grinding and Classification Pro-cess Here we establish the mathematical predictive modelof each unit process in the bauxite grinding and classificationprocess The notations of the indexes decision variables andparameters are listed in Table 5 These notations will be usedfor the model of the grinding and classification process
421 Ball Mill Circuit Model Here 119901119894is the particle per-
centage of 119894th size fraction in the ball mill overflow 119891119895is the
particle percentage of 119894th size fraction in feed ore rate of thefirst classifier recycle 119862 is known and 119864119886
119894is the efficiency of
the first spiral classifier According to a technical report offield investigation and study we have that
119901119894(1 + 119862) =
119889119894119894119891119894+ sum119894minus1
119895=1119894gt1119889119894119895[119864119886119894119901119895(1 + 119862) + 119891
119895]
1 minus 119889119894119894119864119886119894
119889119894119895=
119890119895 119894 = 119895
119894minus1
sum
119896=119895
119888119894119896119888119895119896(119890119896minus 119890119894) 119894 gt 119895
119888119894119895=
minus
119895minus1
sum
119896=119894
119888119894119896119888119895119896 119894 lt 119895
1 119894 = 119895
1
119878119894minus 119878119895
119894minus1
sum
119896=119895
119878119896119887119894119896119888119896119895 119894 gt 119895
119890119895=
1
(1 + 05 sdot 120591119878119895) (1 + 025 sdot 120591119878
119895)2 120591 =
81198751
119872MF (12)
where 120591 is the mean residence time of minerals 1198751is the
internal concentration in ball mill and
1198751=
119872MF (1 + 119862)
119872MF (1 + 119862) +1198821 + 03 (1198821 +1198822) (13)
where119872MF (119905ℎ) is the solid flow of feed ore1198821is the water
addition of the first classifier recycle and1198822is the classifier
water addition 119887119894119895is the breakage distribution function 119878
119894
is the breakage rate function and it satisfied the followingequation
119878119894=
(119886(119889119894119889119900)120572
)
(1 + (119889119894120583)Λ
)
(14)
where 119889119894is the particle with the 119894th size 119889
119900it is a unit when
per millimeter is a unit 119889119900= 1 119889
119894= 119894 (mm) and 119886 120572 120583
and Λ are four key parameters to control the breakage ratefunction
In a concrete grinding and classification process the ballmill size is fixed and the speed of ball mill is constantThrough data acquisition and testing of grinding and clas-sification steady-state loop the regression model between 119886120572 120583 Λ and condition variables size fraction distributioncan be established The input variables are ball filling rate120601119861 solid flow of feed ore 119872MF water addition of the first
classifier recycle1198821 and parameters of feed ore size fraction
distribution 119860MF 119861MF The regression model is
[119886 120572 120583 Λ]119879
=[[
[
11990911
sdot sdot sdot 1199091119895
d
1199091198941
sdot sdot sdot 119909119894119895
]]
]
sdot [1198821 120601119861 119860MF 119861MF 119872MF]119879
(15)
The value of 119909119894119895(119894 = 1 2 3 4 119895 = 1 2 5) can be
obtained by the experimental data regression 119860MF 119861MF can
12 Journal of Applied Mathematics
be obtained from feed ore size fraction distribution and 119865(119894)is the cumulative particle percentage less than the 119894th sizefraction in feed ore and it is represented as follows
119865 (119894) = 1 minus exp (minus119860MF119889119861MF119894
) (16)
422 Spiral Classifier Model 120572119888119894is the particle percentage
of the 119894th size fraction in the first classifier overflow and 119901119894
is the particle percentage of the 119894th size fraction in ball milloverflow The spiral classifier model is as follows
120572119888119894=
119901119894times (1 minus 119864119886
119894)
sum119894=1
(119901119894times (1 minus 119864119886
119894))times 100 (17)
where 119864119886119894is the efficiency of the first spiral classifier and the
mechanism formula of 119864119886119894is shown as follows
119864119886119894= 1 minus exp [minus0693(
119889119894minus 119889min
11988950119888
minus 119889min)
119898
]
+ 119864119886min sdot [1 minus (119889119894minus 119889min119889max
)]
119896
(18)
where 119889119894is the particle with the 119894th size 119889min and 119889max
represent maximum and minimum particle sizes 11988950119888
isthe particle size fraction after correction separation 119898 isseparation accuracy and 119896 is intermix index
Through data acquisition and testing of grinding andclassification steady-state loop the regressionmodel betweenclassification parameters and condition variables size frac-tion distribution can be established The input variablesinclude the solid flow of feed ore 119872MF the classifier wateraddition 119882
2 and the parameters of ball mill overflow size
fraction distribution 119860MF 119861MF The regression model isshown as follows
[119864119886min 11988950119888
119898 119896]119879
=[[
[
11991011
sdot sdot sdot 1199101119895
d
1199101198941
sdot sdot sdot 119910119894119895
]]
]
sdot [119872MF 1198822119860MP 119861MP]
119879
(19)
where the value of 119910119894119895(119894 = 1 2 3 4 119895 = 1 2 3 4) can be
obtained by data regressionThe first-stage overflow 119875
2calculation formula is as
follows
1198752=
119872MF(119872MF +1198821 +1198822)
(20)
Similarly we can get the second spiral classifier model asfollows
1205721198881015840
119894=
120572119888119894times (1 minus 119864119886
1015840
119894)
sum119894(120572119888119894times (1 minus 1198641198861015840))
times 100 (21)
where 1205721198881015840119894is the particle percentage of the 119894th size fraction in
the second classifier overflow 120572119888119894is the particle percentage
of the 119894th size fraction in the first classifier overflow and
1198641198861015840
119894is the efficiency of the second spiral classifier The spiral
classifier model is as follows
1198641198861015840
119894= 1 minus exp[
[
minus0693(1198891015840
119894minus 1198891015840
min1198891015840
50119888minus 1198891015840
min)
1198981015840
]
]
+ 1198641198861015840
min sdot [1 minus (1198891015840
119894minus 1198891015840
min1198891015840max
)]
1198961015840
(22)
where 1198641198861015840min 1198891015840
50119888 1198981015840 and 119896
1015840 are key parameters to theefficiency of the second spiral classifier Through data acqui-sition and testing of grinding and classification steady-stateloop the regression model between classification parametersand condition variables size fraction distribution can beestablished The input variables include solid flow of feedore119872MF and parameters of the first-stage classifier overflowsize fraction distribution 119860
119888 119861119888 which are solved by similar
equation to (20) The regression model is shown as follows
[1198641198861015840
min 1198891015840
5011988811989810158401198961015840]119879
=
[[[
[
1199101015840
11sdot sdot sdot 1199101015840
1119895
d
1199101015840
1198941sdot sdot sdot 119910
1015840
119894119895
]]]
]
sdot [119872MF 119860119888119861119888]119879
(23)
where the value of1199101015840119894119895(119894 = 1 2 3 4 119895 = 1 2 3) can be obtained
by experimental data regression
43 Optimization Model of Grinding and Classification Pro-cess Two objective functions in the process are identifiedone is tomaximize output119891
1(119883) and the other is tomaximize
the small-size fractions (less than 0075mm fractions) in thesecond-stage overflow 119891
2(119883) It is also necessary to ensure
that the grinding productmeets all of other technical require-ments and the least disturbance in the following flotationcircuit As the constraints the feed load of the grinding circuit119872MF the steel ball filling rate 120601
119861 the first and the second
overflows 1198751and 119875
2 and the particle percentage of fine size
fraction in the first and the second classifier overflows 120572119888minus0075
and 1205721198881015840minus0075
should be within the user specified boundsThe operation variables are the solid flow of feed ore
119872MF water addition of the first classifier recycle 1198821 ball
filling rate 120601119861 and water addition of the second classifier119882
2
Based on all of the above grinding and classification processmultiobjective optimization model is as follows
max 119865 = max [1198911(119883) 119891
2(119883)]
1198911(119883) = 119872MF
1198912(119883) = 120572119888
1015840
minus0075
= 119891 (119872MF11988211198822 120601119861)
st 119872MFmin le 119872MF le 119872MFmax120601119861min le 120601119861 le 120601119861max1198751min le 1198751 le 1198751max1198752min le 1198752 le 1198752max120572119888minus0075
ge 120572119888min
1205721198881015840
minus0075ge 1205721198881015840
min
(24)
Journal of Applied Mathematics 13
Table 6 The optimization results calculated by the proposed algo-rithm
Number 1198911(119883) (th) 119891
2(119883) ()
1 92972 908142 91810 919003 92360 909234 90620 934605 91310 924946 90170 938007 89390 952008 89480 949329 91530 9240010 89298 9576311 89824 9454912 89800 9439513 89710 9461814 91170 9280015 91880 9184016 92190 9128117 89870 9409018 91000 9328519 91960 9152020 91124 93221
85
87
89
91
93
95
97
74000 79000 84000 94000
Optimization resultsThe raw data
f1(x)
f2(x)
()
89000
Figure 8 The comparison chart between optimization results andindustrial data
In (24) 119891(119872MF11988211198822 120601119861) implicates the model ofgrinding and classification represented by (12)ndash(23) Theproposed algorithm is applied to solve the problem and theoptimization results are shown in Table 6
With the practical process data from a grinding circuitof a mineral plant the simulation of this hybrid intelligentmethod adopted the same parameters on the variation infresh slurry feed velocity density particle size distributionand cyclone feed operating configurations
The comparison of production data and optimizationresults in Table 6 is shown in Figure 8 where ldquoXrdquo repre-sents the proposed algorithm optimization results and ldquoIrdquorepresents the original data collected from the field withoutoptimization of raw data According to the objectives thedata point closer to the upper right edge is more beneficialObviously the proposed optimization result is far betterthan the original data indicating the effectiveness of theoptimization approach
44 TOPSIS Method for Solution Selection The resolution ofamultiobjective optimization problemdoes not endwhen thePareto-optimal set is found In practical operational prob-lems a single solution must be selected TOPSIS [36] is auseful technique in dealing with multiattribute or multicrite-ria decision-making (MADMMCDM) problems in the realworld The standard TOPSIS method attempts to choosealternatives that simultaneously have the shortest distancefrom the positive-ideal solution and the farthest distancefrom the negative-ideal solution According to the TOPSISmethod the relative closeness coefficient is calculated andthe best solution in Table 6 is the solution number 10 as1198911(119883) = 89298 (119905ℎ) and 119891
2(119883) = 95763 () The
corresponding decision variables are 119872MF = 89298 (119905ℎ)120601119861= 32119882
1= 75127 (119905ℎ) and119882
2= 16296 (119905ℎ)
5 Conclusions
Promoted by the requirements of engineering optimizationin complex practical processes of grinding and classificationwe proposed a hybrid multiobjective differential evolutionalgorithm with a few beneficial features integrated Firstly anarchiving mechanism for infeasible solutions is establishedwith functional partitioned multi-population which aims todirect the population to approach or land in the feasibleregion from different directions during evolution Secondlywe propose an infeasible constraint violations function toselect infeasible population with better performance sothat they are allowed to be saved and to participate inthe subsequent evolution Thirdly a nondominating rankingstrategy is designed to improve the crowding-distance sortingand return uniform distribution of Pareto solutions Finallythe simplex method is inserted in the differential evolutionprocess to purposefully enrich the diversity without excessivecomputation cost The advantage of the proposed algorithmis the exemption from constructing penalty function andthe preservation of meaningful infeasible solutions directlySimulation results on benchmarks indicate that the proposedalgorithm can converge quickly and effectively to the truePareto frontier with better distribution
Based on the investigated information about grinding cir-cuit process we established a multiobjective optimal modelwith equations frommechanism knowledge parameters rec-ognized by data regression and constraints of technicalrequirements The nonlinear multiobjective optimizationmodel is too complicated to be solved by traditional gradient-based algorithmsThe proposed hybrid differential evolutionalgorithm is applied and tested to achieve a Pareto solution
14 Journal of Applied Mathematics
set It is proven to be valuable for operation decision makingin the industrial process and showed superiority to the oper-ation carried out in the production In fact many operatingparameters in complex processes are highly coupled andconflicting with each other The optimal operation of theentire production process is very difficult to obtain bymanualcalculation let alone the fluctuation situation of processconditions The application case indicates that the proposedmethod has good performance and is helpful to inspirefurther research on evolutionary methods for engineeringoptimization
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgments
This work is supported by the National Natural ScienceFoundation of China (61374156 61273187 and 61134006)the National Key Technology Research and DevelopmentProgram of the Ministry of Science and Technology ofChina (2012BAK09B04) the Science Fund for CreativeResearch Groups of the National Natural Science Foundationof China (61321003) the Fund for Doctor Station of theMinistry of Education (20110162130011 and 20100162120019)and theHunanProvince Science andTechnology PlanProject(2012CK4018)
References
[1] K Mitra and R Gopinath ldquoMultiobjective optimization ofan industrial grinding operation using elitist nondominatedsorting genetic algorithmrdquo Chemical Engineering Science vol59 no 2 pp 385ndash396 2004
[2] G Yu T Y Chai and X C Luo ldquoMultiobjective productionplanning optimization using hybrid evolutionary algorithmsfor mineral processingrdquo IEEE Transactions on EvolutionaryComputation vol 15 no 4 pp 487ndash514 2011
[3] T Y Ma W H Gui Y L Wang and C H Yang ldquoDynamicoptimization control for grinding and classification processrdquoControl andDecision vol 27 no 2 pp 286ndash290 2012 (Chinese)
[4] F Y Edgeworth Mathematical Physics An Essay on the Appli-cation of Mathematics to the Moral Sciences C Kegan Paul ampCompany London UK 1881
[5] V Pareto Cours drsquoEconomie Politique F Rouge LausanneSwitzerland 1896
[6] X F Chen W H Gui Y L Wang and L Cen ldquoMulti-stepoptimal control of complex process a genetic programmingstrategy and its applicationrdquo Engineering Applications of Artifi-cial Intelligence vol 17 no 5 pp 491ndash500 2004
[7] C A C Coello ldquoEvolutionary multi-objective optimizationa historical view of the fieldrdquo IEEE Computational IntelligenceMagazine vol 1 no 1 pp 28ndash36 2006
[8] K Deb A Pratap S Agarwal and T Meyarivan ldquoA fastand elitist multiobjective genetic algorithm NSGA-IIrdquo IEEETransactions on Evolutionary Computation vol 6 no 2 pp 182ndash197 2002
[9] C M Fonesca and P J Fleming ldquoGenetic algorithms formulti objective optimization formulation discussion and gen-eralizationrdquo in Proceedings of the 5th International Conferenceon Genetic Algorithms pp 415ndash423 Morgan Kaufmann SanMateo Calif USA 1993
[10] S Sengupta S Das M Nasir A V Vasilakos and W PedryczldquoEnergy-efficient differentiated coverage of dynamic objectsusing an improved evolutionary multi-objective optimizationalgorithm with fuzzy-dominancerdquo in Proceedings of IEEECongress on Evolutionary Computation pp 1ndash8 IEEE PressBrisbane Australia 2012
[11] K V Price R M Storn and J A Lampinen DifferentialEvolution a Practical Approach to Global Optimization SpringerBerlin Germany 2005
[12] W Y Gong and Z H Cai ldquoAn improved multiobjective dif-ferential evolution based on Pareto-adaptive 120576-dominance andorthogonal designrdquo European Journal of Operational Researchvol 198 no 2 pp 576ndash601 2009
[13] X H Zeng W K Wong and S Y Leung ldquoAn operatorallocation optimization model for balancing control of thehybrid assembly lines using Pareto utility discrete differentialevolution algorithmrdquo Computers and Operations Research vol39 no 5 pp 1145ndash1159 2012
[14] E Zitzler and L Thiele ldquoMultiobjective evolutionary algo-rithms a comparative case study and the strength Paretoapproachrdquo IEEETransactions onEvolutionaryComputation vol3 no 4 pp 257ndash271 1999
[15] A Al-Ani A Alsukker and R N Khushaba ldquoFeature subsetselection using differential evolution and a wheel based searchstrategyrdquo Swarm and Evolutionary Computation vol 9 pp 15ndash26 2013
[16] G Y Li and M G Liu ldquoThe summary of differential evolutionalgorithm and its improvementsrdquo in Proceedings of the 3rdInternational Conference on Advanced Computer Theory andEngineering (ICACTE rsquo10) pp V3153ndashV3156 Chengdu ChinaAugust 2010
[17] H A Abbass ldquoThe self-adaptive Pareto differential evolutionalgorithmrdquo in Proceedings of the IEEE Congress on EvolutionaryComputation pp 831ndash836 IEEE Press Honolulu Hawaii USAMay 2002
[18] N K Madavan ldquoMultiobjective optimization using a Paretodifferential evolution approachrdquo in Proceedings of the IEEECongress on Evolutionary Computation pp 1145ndash1150 IEEEPress Honolulu Hawaii USA May 2002
[19] F Xue A C Sanderson and R J Graves ldquoPareto-based multi-objective differential evolutionrdquo in Proceedings of the IEEECongress on Evolutionary Computation pp 862ndash869 IEEEPress Piscataway NJ USA December 2003
[20] T Robic and B Filipi ldquoDEMO differential evolution formultiobjective optimizationrdquo in Evolutionary Multi-CriterionOptimization pp 520ndash533 Springer Berlin Germany 2005
[21] V L Huang A K Qin P N Suganthan and M F TasgetirenldquoMulti-objective optimization based on self-adaptive differen-tial evolution algorithmrdquo in Proceedings of the IEEE Congresson Evolutionary Computation (CEC rsquo07) pp 3601ndash3608 IEEEPress Singapore September 2007
[22] V L Huang S Z Zhao R Mallipeddi and P N SuganthanldquoMulti-objective optimization using self-adaptive differentialevolution algorithmrdquo in Proceedings of the IEEE Congress onEvolutionary Computation (CEC rsquo09) pp 190ndash194 IEEE PressTrondheim Norway May 2009
Journal of Applied Mathematics 15
[23] J A Adeyemo and F A O Otieno ldquoMulti-objective differentialevolution algorithm for solving engineering problemsrdquo Journalof Applied Sciences vol 9 no 20 pp 3652ndash3661 2009
[24] Y Wang and Z Cai ldquoCombining multiobjective optimizationwith differential evolution to solve constrained optimizationproblemsrdquo IEEE Transactions on Evolutionary Computationvol 16 no 1 pp 117ndash134 2012
[25] A Homaifar C X Qi and S H Lai ldquoConstrained optimizationvia genetic algorithmsrdquo Simulation vol 62 no 4 pp 242ndash2531994
[26] J A Joines and C R Houck ldquoOn the use of non-stationarypenalty functions to solve nonlinear constrained optimizationproblems with GArsquosrdquo in Proceedings of the 1st IEEE Conferenceon Evolutionary Computation pp 579ndash584 Orlando Fla USAJune 1994
[27] F Jimenez and J L Verdegay ldquoEvolutionary techniques forconstrained optimization problemsrdquo in Proceedings of the 7thEuropean Congress on Intelligent Techniques and Soft Computing(EUFIT rsquo99) Aachen Germany 1999
[28] C Sun J Zeng and J Pan ldquoAn improved vector particleswarm optimization for constrained optimization problemsrdquoInformation Sciences vol 181 no 6 pp 1153ndash1163 2011
[29] Z Michalewicz and M Schoenauer ldquoEvolutionary algorithmsfor constrained parameter optimization problemsrdquo Evolution-ary Computation vol 4 no 1 pp 1ndash32 1996
[30] C A C Coello and A D Christiansen ldquoMoses a multiob-jective optimization tool for engineering designrdquo EngineeringOptimization vol 31 no 1ndash3 pp 337ndash368 1999
[31] Y G Woldesenbet G G Yen and B G Tessema ldquoConstrainthandling in multiobjective evolutionary optimizationrdquo IEEETransactions on Evolutionary Computation vol 13 no 3 pp514ndash525 2009
[32] J A Nelder and R Mead ldquoA simplex method for functionminimizationrdquoThe Computer Journal vol 7 no 4 pp 308ndash3131965
[33] D A van Veldhuizen and G B Lamont ldquoMultiobjective evolu-tionary algorithms analyzing the state-of-the-artrdquo EvolutionaryComputation vol 8 no 2 pp 125ndash147 2000
[34] R Farmani D A Savic andG AWalters ldquoEvolutionarymulti-objective optimization in water distribution network designrdquoEngineering Optimization vol 37 no 2 pp 167ndash183 2005
[35] K Deb and T Goel Controlled Elitist Non-Dominated SortingGenetic Algorithms for Better Convergence Springer BerlinGermany 2001
[36] Y J Lai T Y Liu and C L Hwang ldquoTopsis for MODMrdquoEuropean Journal of Operational Research vol 76 no 3 pp486ndash500 1994
Submit your manuscripts athttpwwwhindawicom
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical Problems in Engineering
Hindawi Publishing Corporationhttpwwwhindawicom
Differential EquationsInternational Journal of
Volume 2014
Applied MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Probability and StatisticsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical PhysicsAdvances in
Complex AnalysisJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
OptimizationJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
CombinatoricsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Operations ResearchAdvances in
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Function Spaces
Abstract and Applied AnalysisHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of Mathematics and Mathematical Sciences
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Algebra
Discrete Dynamics in Nature and Society
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Decision SciencesAdvances in
Discrete MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014 Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Stochastic AnalysisInternational Journal of
10 Journal of Applied Mathematics
15
10
05
00 05 10 15
Pareto frontierProposed algorithm
f2
f1
(g) TNK calculated by the proposed algorithm
Pareto frontier
15
10
05
00 05 10 15
f2
f1
CNSGA-II
(h) TNK calculated by CNSGA-II
Figure 6 The comparison chart of Pareto frontier
Table 4 The comparison of performance
Test function Algorithm 120574 SP
CTP1 CNSGA-Π 0021317 plusmn 0000323 0873321 plusmn 008725
The proposed 0009836 plusmn 0000410 0567933 plusmn 001845
CTP2 CNSGA-Π 0011120 plusmn 0000753 078314 plusmn 002843
The proposed 0007013 plusmn 0000554 0296543 plusmn 000453
BNH CNSGA-Π 0014947 plusmn 0000632 0336941 plusmn 000917
The proposed 0013766 plusmn 0000043 0209321 plusmn 000561
TNK CNSGA-Π 0013235 plusmn 0000740 0464542 plusmn 000730
The proposed 0006435 plusmn 0000017 0224560 plusmn 000159
Feed water Water
Feed ore Ball mill OverflowMixing
PumpClassifierwater
additionFirst
classifierrecycle First-
overflow Spiral classifier 2Second-stage
to flotationSpiral classifier 1 Second classifier recycle(rough concentrate)
addition
pump
stageoverflow
Figure 7 Flow diagram of the grinding and classification process
algorithms based on the two criterions 120574 and SP as shownin Table 4 It can be observed from the data in Table 4that the proposed algorithm performs significantly betterthan the classical CNSGA-II algorithm in convergence anddistribution uniformityThe simulation results show that thisalgorithm can accurately converge to global Pareto solutionsand can maintain diversity of population
4 Optimization of Grinding andClassification Process
41 Bauxite Grinding and Classification Process Thegrindingand classification process is the key preparation for thebauxite mineral processing Here we consider a bauxitegrinding process in a certain mineral company with singlegrinding and two-stage classification as shown in Figure 7
The process consists in a grinding ball mill and twospiral classifiers First classifier recycle will be put back tothe ball mill for regrinding and the first-stage overflowwill be put into second spiral classifier after being mixedwith water the second classifier recycle will be prepared forBayer production as the rough concentrate and the second-stage overflow will be sent to the next flotation processThe production objectives are composed of the productionyields technically represented by the solid flow of feed oresince the process is nonstorable and the mineral processingquality represented by percentage of the small-size fractionsof mineral particles in the second-stage overflows
Journal of Applied Mathematics 11
Table 5 Notations for the model of the grinding and classificationprocess
Notation Description
119901119894
Particle percentage of the 119894th size fraction inthe ball mill overflow
119891119895
Particle percentage of the 119894th size fraction inthe Feed ore
119862 Rate of the first classifier recycle119864119886119894 The efficiency of the first spiral classifier
120591 Themean residence time1198751 The internal concentration in ball mill119872MF The solid flow of feed ore1198821 The water addition of the first classifier recycle
1198822 The classifier water addition
119887119894119895 The breakage distribution function119878119894 The breakage rate function119889119894 The particle with the 119894th size
120572119888119894
The particle percentage of the 119894th size in thefirst classifier overflow
119889119900 The unit size of the particle1198752 First-stage overflow120601119861 Ball filling rate
119860119888 119861119888
Parameters of the first-stage classifier overflowsize fraction distribution
119864119886119894 The efficiency of the first spiral classifier
11988950119888
The particle size fraction after correctionseparation
119898 The separation accuracy119896 The intermix index
1198641198861015840
min 1198891015840
50119888 1198981015840 1198961015840 The corresponding key parameters to theefficiency of the second spiral classifier
119886 120572 120583 ΛFour key parameters to control the breakagerate function
119889min 119889max Theminimum and maximum particle sizes119860MF 119861MF Parameters of feed ore size fraction distribution
119865(119894)The cumulative particle percentage less thanthe 119894th size fraction in feed ore
1205721198881015840
119894
The particle percentage of the 119894th size fractionin the second classifier overflow
1198641198861015840
119894 The efficiency of the second spiral classifier
42 Predictive Model of the Grinding and Classification Pro-cess Here we establish the mathematical predictive modelof each unit process in the bauxite grinding and classificationprocess The notations of the indexes decision variables andparameters are listed in Table 5 These notations will be usedfor the model of the grinding and classification process
421 Ball Mill Circuit Model Here 119901119894is the particle per-
centage of 119894th size fraction in the ball mill overflow 119891119895is the
particle percentage of 119894th size fraction in feed ore rate of thefirst classifier recycle 119862 is known and 119864119886
119894is the efficiency of
the first spiral classifier According to a technical report offield investigation and study we have that
119901119894(1 + 119862) =
119889119894119894119891119894+ sum119894minus1
119895=1119894gt1119889119894119895[119864119886119894119901119895(1 + 119862) + 119891
119895]
1 minus 119889119894119894119864119886119894
119889119894119895=
119890119895 119894 = 119895
119894minus1
sum
119896=119895
119888119894119896119888119895119896(119890119896minus 119890119894) 119894 gt 119895
119888119894119895=
minus
119895minus1
sum
119896=119894
119888119894119896119888119895119896 119894 lt 119895
1 119894 = 119895
1
119878119894minus 119878119895
119894minus1
sum
119896=119895
119878119896119887119894119896119888119896119895 119894 gt 119895
119890119895=
1
(1 + 05 sdot 120591119878119895) (1 + 025 sdot 120591119878
119895)2 120591 =
81198751
119872MF (12)
where 120591 is the mean residence time of minerals 1198751is the
internal concentration in ball mill and
1198751=
119872MF (1 + 119862)
119872MF (1 + 119862) +1198821 + 03 (1198821 +1198822) (13)
where119872MF (119905ℎ) is the solid flow of feed ore1198821is the water
addition of the first classifier recycle and1198822is the classifier
water addition 119887119894119895is the breakage distribution function 119878
119894
is the breakage rate function and it satisfied the followingequation
119878119894=
(119886(119889119894119889119900)120572
)
(1 + (119889119894120583)Λ
)
(14)
where 119889119894is the particle with the 119894th size 119889
119900it is a unit when
per millimeter is a unit 119889119900= 1 119889
119894= 119894 (mm) and 119886 120572 120583
and Λ are four key parameters to control the breakage ratefunction
In a concrete grinding and classification process the ballmill size is fixed and the speed of ball mill is constantThrough data acquisition and testing of grinding and clas-sification steady-state loop the regression model between 119886120572 120583 Λ and condition variables size fraction distributioncan be established The input variables are ball filling rate120601119861 solid flow of feed ore 119872MF water addition of the first
classifier recycle1198821 and parameters of feed ore size fraction
distribution 119860MF 119861MF The regression model is
[119886 120572 120583 Λ]119879
=[[
[
11990911
sdot sdot sdot 1199091119895
d
1199091198941
sdot sdot sdot 119909119894119895
]]
]
sdot [1198821 120601119861 119860MF 119861MF 119872MF]119879
(15)
The value of 119909119894119895(119894 = 1 2 3 4 119895 = 1 2 5) can be
obtained by the experimental data regression 119860MF 119861MF can
12 Journal of Applied Mathematics
be obtained from feed ore size fraction distribution and 119865(119894)is the cumulative particle percentage less than the 119894th sizefraction in feed ore and it is represented as follows
119865 (119894) = 1 minus exp (minus119860MF119889119861MF119894
) (16)
422 Spiral Classifier Model 120572119888119894is the particle percentage
of the 119894th size fraction in the first classifier overflow and 119901119894
is the particle percentage of the 119894th size fraction in ball milloverflow The spiral classifier model is as follows
120572119888119894=
119901119894times (1 minus 119864119886
119894)
sum119894=1
(119901119894times (1 minus 119864119886
119894))times 100 (17)
where 119864119886119894is the efficiency of the first spiral classifier and the
mechanism formula of 119864119886119894is shown as follows
119864119886119894= 1 minus exp [minus0693(
119889119894minus 119889min
11988950119888
minus 119889min)
119898
]
+ 119864119886min sdot [1 minus (119889119894minus 119889min119889max
)]
119896
(18)
where 119889119894is the particle with the 119894th size 119889min and 119889max
represent maximum and minimum particle sizes 11988950119888
isthe particle size fraction after correction separation 119898 isseparation accuracy and 119896 is intermix index
Through data acquisition and testing of grinding andclassification steady-state loop the regressionmodel betweenclassification parameters and condition variables size frac-tion distribution can be established The input variablesinclude the solid flow of feed ore 119872MF the classifier wateraddition 119882
2 and the parameters of ball mill overflow size
fraction distribution 119860MF 119861MF The regression model isshown as follows
[119864119886min 11988950119888
119898 119896]119879
=[[
[
11991011
sdot sdot sdot 1199101119895
d
1199101198941
sdot sdot sdot 119910119894119895
]]
]
sdot [119872MF 1198822119860MP 119861MP]
119879
(19)
where the value of 119910119894119895(119894 = 1 2 3 4 119895 = 1 2 3 4) can be
obtained by data regressionThe first-stage overflow 119875
2calculation formula is as
follows
1198752=
119872MF(119872MF +1198821 +1198822)
(20)
Similarly we can get the second spiral classifier model asfollows
1205721198881015840
119894=
120572119888119894times (1 minus 119864119886
1015840
119894)
sum119894(120572119888119894times (1 minus 1198641198861015840))
times 100 (21)
where 1205721198881015840119894is the particle percentage of the 119894th size fraction in
the second classifier overflow 120572119888119894is the particle percentage
of the 119894th size fraction in the first classifier overflow and
1198641198861015840
119894is the efficiency of the second spiral classifier The spiral
classifier model is as follows
1198641198861015840
119894= 1 minus exp[
[
minus0693(1198891015840
119894minus 1198891015840
min1198891015840
50119888minus 1198891015840
min)
1198981015840
]
]
+ 1198641198861015840
min sdot [1 minus (1198891015840
119894minus 1198891015840
min1198891015840max
)]
1198961015840
(22)
where 1198641198861015840min 1198891015840
50119888 1198981015840 and 119896
1015840 are key parameters to theefficiency of the second spiral classifier Through data acqui-sition and testing of grinding and classification steady-stateloop the regression model between classification parametersand condition variables size fraction distribution can beestablished The input variables include solid flow of feedore119872MF and parameters of the first-stage classifier overflowsize fraction distribution 119860
119888 119861119888 which are solved by similar
equation to (20) The regression model is shown as follows
[1198641198861015840
min 1198891015840
5011988811989810158401198961015840]119879
=
[[[
[
1199101015840
11sdot sdot sdot 1199101015840
1119895
d
1199101015840
1198941sdot sdot sdot 119910
1015840
119894119895
]]]
]
sdot [119872MF 119860119888119861119888]119879
(23)
where the value of1199101015840119894119895(119894 = 1 2 3 4 119895 = 1 2 3) can be obtained
by experimental data regression
43 Optimization Model of Grinding and Classification Pro-cess Two objective functions in the process are identifiedone is tomaximize output119891
1(119883) and the other is tomaximize
the small-size fractions (less than 0075mm fractions) in thesecond-stage overflow 119891
2(119883) It is also necessary to ensure
that the grinding productmeets all of other technical require-ments and the least disturbance in the following flotationcircuit As the constraints the feed load of the grinding circuit119872MF the steel ball filling rate 120601
119861 the first and the second
overflows 1198751and 119875
2 and the particle percentage of fine size
fraction in the first and the second classifier overflows 120572119888minus0075
and 1205721198881015840minus0075
should be within the user specified boundsThe operation variables are the solid flow of feed ore
119872MF water addition of the first classifier recycle 1198821 ball
filling rate 120601119861 and water addition of the second classifier119882
2
Based on all of the above grinding and classification processmultiobjective optimization model is as follows
max 119865 = max [1198911(119883) 119891
2(119883)]
1198911(119883) = 119872MF
1198912(119883) = 120572119888
1015840
minus0075
= 119891 (119872MF11988211198822 120601119861)
st 119872MFmin le 119872MF le 119872MFmax120601119861min le 120601119861 le 120601119861max1198751min le 1198751 le 1198751max1198752min le 1198752 le 1198752max120572119888minus0075
ge 120572119888min
1205721198881015840
minus0075ge 1205721198881015840
min
(24)
Journal of Applied Mathematics 13
Table 6 The optimization results calculated by the proposed algo-rithm
Number 1198911(119883) (th) 119891
2(119883) ()
1 92972 908142 91810 919003 92360 909234 90620 934605 91310 924946 90170 938007 89390 952008 89480 949329 91530 9240010 89298 9576311 89824 9454912 89800 9439513 89710 9461814 91170 9280015 91880 9184016 92190 9128117 89870 9409018 91000 9328519 91960 9152020 91124 93221
85
87
89
91
93
95
97
74000 79000 84000 94000
Optimization resultsThe raw data
f1(x)
f2(x)
()
89000
Figure 8 The comparison chart between optimization results andindustrial data
In (24) 119891(119872MF11988211198822 120601119861) implicates the model ofgrinding and classification represented by (12)ndash(23) Theproposed algorithm is applied to solve the problem and theoptimization results are shown in Table 6
With the practical process data from a grinding circuitof a mineral plant the simulation of this hybrid intelligentmethod adopted the same parameters on the variation infresh slurry feed velocity density particle size distributionand cyclone feed operating configurations
The comparison of production data and optimizationresults in Table 6 is shown in Figure 8 where ldquoXrdquo repre-sents the proposed algorithm optimization results and ldquoIrdquorepresents the original data collected from the field withoutoptimization of raw data According to the objectives thedata point closer to the upper right edge is more beneficialObviously the proposed optimization result is far betterthan the original data indicating the effectiveness of theoptimization approach
44 TOPSIS Method for Solution Selection The resolution ofamultiobjective optimization problemdoes not endwhen thePareto-optimal set is found In practical operational prob-lems a single solution must be selected TOPSIS [36] is auseful technique in dealing with multiattribute or multicrite-ria decision-making (MADMMCDM) problems in the realworld The standard TOPSIS method attempts to choosealternatives that simultaneously have the shortest distancefrom the positive-ideal solution and the farthest distancefrom the negative-ideal solution According to the TOPSISmethod the relative closeness coefficient is calculated andthe best solution in Table 6 is the solution number 10 as1198911(119883) = 89298 (119905ℎ) and 119891
2(119883) = 95763 () The
corresponding decision variables are 119872MF = 89298 (119905ℎ)120601119861= 32119882
1= 75127 (119905ℎ) and119882
2= 16296 (119905ℎ)
5 Conclusions
Promoted by the requirements of engineering optimizationin complex practical processes of grinding and classificationwe proposed a hybrid multiobjective differential evolutionalgorithm with a few beneficial features integrated Firstly anarchiving mechanism for infeasible solutions is establishedwith functional partitioned multi-population which aims todirect the population to approach or land in the feasibleregion from different directions during evolution Secondlywe propose an infeasible constraint violations function toselect infeasible population with better performance sothat they are allowed to be saved and to participate inthe subsequent evolution Thirdly a nondominating rankingstrategy is designed to improve the crowding-distance sortingand return uniform distribution of Pareto solutions Finallythe simplex method is inserted in the differential evolutionprocess to purposefully enrich the diversity without excessivecomputation cost The advantage of the proposed algorithmis the exemption from constructing penalty function andthe preservation of meaningful infeasible solutions directlySimulation results on benchmarks indicate that the proposedalgorithm can converge quickly and effectively to the truePareto frontier with better distribution
Based on the investigated information about grinding cir-cuit process we established a multiobjective optimal modelwith equations frommechanism knowledge parameters rec-ognized by data regression and constraints of technicalrequirements The nonlinear multiobjective optimizationmodel is too complicated to be solved by traditional gradient-based algorithmsThe proposed hybrid differential evolutionalgorithm is applied and tested to achieve a Pareto solution
14 Journal of Applied Mathematics
set It is proven to be valuable for operation decision makingin the industrial process and showed superiority to the oper-ation carried out in the production In fact many operatingparameters in complex processes are highly coupled andconflicting with each other The optimal operation of theentire production process is very difficult to obtain bymanualcalculation let alone the fluctuation situation of processconditions The application case indicates that the proposedmethod has good performance and is helpful to inspirefurther research on evolutionary methods for engineeringoptimization
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgments
This work is supported by the National Natural ScienceFoundation of China (61374156 61273187 and 61134006)the National Key Technology Research and DevelopmentProgram of the Ministry of Science and Technology ofChina (2012BAK09B04) the Science Fund for CreativeResearch Groups of the National Natural Science Foundationof China (61321003) the Fund for Doctor Station of theMinistry of Education (20110162130011 and 20100162120019)and theHunanProvince Science andTechnology PlanProject(2012CK4018)
References
[1] K Mitra and R Gopinath ldquoMultiobjective optimization ofan industrial grinding operation using elitist nondominatedsorting genetic algorithmrdquo Chemical Engineering Science vol59 no 2 pp 385ndash396 2004
[2] G Yu T Y Chai and X C Luo ldquoMultiobjective productionplanning optimization using hybrid evolutionary algorithmsfor mineral processingrdquo IEEE Transactions on EvolutionaryComputation vol 15 no 4 pp 487ndash514 2011
[3] T Y Ma W H Gui Y L Wang and C H Yang ldquoDynamicoptimization control for grinding and classification processrdquoControl andDecision vol 27 no 2 pp 286ndash290 2012 (Chinese)
[4] F Y Edgeworth Mathematical Physics An Essay on the Appli-cation of Mathematics to the Moral Sciences C Kegan Paul ampCompany London UK 1881
[5] V Pareto Cours drsquoEconomie Politique F Rouge LausanneSwitzerland 1896
[6] X F Chen W H Gui Y L Wang and L Cen ldquoMulti-stepoptimal control of complex process a genetic programmingstrategy and its applicationrdquo Engineering Applications of Artifi-cial Intelligence vol 17 no 5 pp 491ndash500 2004
[7] C A C Coello ldquoEvolutionary multi-objective optimizationa historical view of the fieldrdquo IEEE Computational IntelligenceMagazine vol 1 no 1 pp 28ndash36 2006
[8] K Deb A Pratap S Agarwal and T Meyarivan ldquoA fastand elitist multiobjective genetic algorithm NSGA-IIrdquo IEEETransactions on Evolutionary Computation vol 6 no 2 pp 182ndash197 2002
[9] C M Fonesca and P J Fleming ldquoGenetic algorithms formulti objective optimization formulation discussion and gen-eralizationrdquo in Proceedings of the 5th International Conferenceon Genetic Algorithms pp 415ndash423 Morgan Kaufmann SanMateo Calif USA 1993
[10] S Sengupta S Das M Nasir A V Vasilakos and W PedryczldquoEnergy-efficient differentiated coverage of dynamic objectsusing an improved evolutionary multi-objective optimizationalgorithm with fuzzy-dominancerdquo in Proceedings of IEEECongress on Evolutionary Computation pp 1ndash8 IEEE PressBrisbane Australia 2012
[11] K V Price R M Storn and J A Lampinen DifferentialEvolution a Practical Approach to Global Optimization SpringerBerlin Germany 2005
[12] W Y Gong and Z H Cai ldquoAn improved multiobjective dif-ferential evolution based on Pareto-adaptive 120576-dominance andorthogonal designrdquo European Journal of Operational Researchvol 198 no 2 pp 576ndash601 2009
[13] X H Zeng W K Wong and S Y Leung ldquoAn operatorallocation optimization model for balancing control of thehybrid assembly lines using Pareto utility discrete differentialevolution algorithmrdquo Computers and Operations Research vol39 no 5 pp 1145ndash1159 2012
[14] E Zitzler and L Thiele ldquoMultiobjective evolutionary algo-rithms a comparative case study and the strength Paretoapproachrdquo IEEETransactions onEvolutionaryComputation vol3 no 4 pp 257ndash271 1999
[15] A Al-Ani A Alsukker and R N Khushaba ldquoFeature subsetselection using differential evolution and a wheel based searchstrategyrdquo Swarm and Evolutionary Computation vol 9 pp 15ndash26 2013
[16] G Y Li and M G Liu ldquoThe summary of differential evolutionalgorithm and its improvementsrdquo in Proceedings of the 3rdInternational Conference on Advanced Computer Theory andEngineering (ICACTE rsquo10) pp V3153ndashV3156 Chengdu ChinaAugust 2010
[17] H A Abbass ldquoThe self-adaptive Pareto differential evolutionalgorithmrdquo in Proceedings of the IEEE Congress on EvolutionaryComputation pp 831ndash836 IEEE Press Honolulu Hawaii USAMay 2002
[18] N K Madavan ldquoMultiobjective optimization using a Paretodifferential evolution approachrdquo in Proceedings of the IEEECongress on Evolutionary Computation pp 1145ndash1150 IEEEPress Honolulu Hawaii USA May 2002
[19] F Xue A C Sanderson and R J Graves ldquoPareto-based multi-objective differential evolutionrdquo in Proceedings of the IEEECongress on Evolutionary Computation pp 862ndash869 IEEEPress Piscataway NJ USA December 2003
[20] T Robic and B Filipi ldquoDEMO differential evolution formultiobjective optimizationrdquo in Evolutionary Multi-CriterionOptimization pp 520ndash533 Springer Berlin Germany 2005
[21] V L Huang A K Qin P N Suganthan and M F TasgetirenldquoMulti-objective optimization based on self-adaptive differen-tial evolution algorithmrdquo in Proceedings of the IEEE Congresson Evolutionary Computation (CEC rsquo07) pp 3601ndash3608 IEEEPress Singapore September 2007
[22] V L Huang S Z Zhao R Mallipeddi and P N SuganthanldquoMulti-objective optimization using self-adaptive differentialevolution algorithmrdquo in Proceedings of the IEEE Congress onEvolutionary Computation (CEC rsquo09) pp 190ndash194 IEEE PressTrondheim Norway May 2009
Journal of Applied Mathematics 15
[23] J A Adeyemo and F A O Otieno ldquoMulti-objective differentialevolution algorithm for solving engineering problemsrdquo Journalof Applied Sciences vol 9 no 20 pp 3652ndash3661 2009
[24] Y Wang and Z Cai ldquoCombining multiobjective optimizationwith differential evolution to solve constrained optimizationproblemsrdquo IEEE Transactions on Evolutionary Computationvol 16 no 1 pp 117ndash134 2012
[25] A Homaifar C X Qi and S H Lai ldquoConstrained optimizationvia genetic algorithmsrdquo Simulation vol 62 no 4 pp 242ndash2531994
[26] J A Joines and C R Houck ldquoOn the use of non-stationarypenalty functions to solve nonlinear constrained optimizationproblems with GArsquosrdquo in Proceedings of the 1st IEEE Conferenceon Evolutionary Computation pp 579ndash584 Orlando Fla USAJune 1994
[27] F Jimenez and J L Verdegay ldquoEvolutionary techniques forconstrained optimization problemsrdquo in Proceedings of the 7thEuropean Congress on Intelligent Techniques and Soft Computing(EUFIT rsquo99) Aachen Germany 1999
[28] C Sun J Zeng and J Pan ldquoAn improved vector particleswarm optimization for constrained optimization problemsrdquoInformation Sciences vol 181 no 6 pp 1153ndash1163 2011
[29] Z Michalewicz and M Schoenauer ldquoEvolutionary algorithmsfor constrained parameter optimization problemsrdquo Evolution-ary Computation vol 4 no 1 pp 1ndash32 1996
[30] C A C Coello and A D Christiansen ldquoMoses a multiob-jective optimization tool for engineering designrdquo EngineeringOptimization vol 31 no 1ndash3 pp 337ndash368 1999
[31] Y G Woldesenbet G G Yen and B G Tessema ldquoConstrainthandling in multiobjective evolutionary optimizationrdquo IEEETransactions on Evolutionary Computation vol 13 no 3 pp514ndash525 2009
[32] J A Nelder and R Mead ldquoA simplex method for functionminimizationrdquoThe Computer Journal vol 7 no 4 pp 308ndash3131965
[33] D A van Veldhuizen and G B Lamont ldquoMultiobjective evolu-tionary algorithms analyzing the state-of-the-artrdquo EvolutionaryComputation vol 8 no 2 pp 125ndash147 2000
[34] R Farmani D A Savic andG AWalters ldquoEvolutionarymulti-objective optimization in water distribution network designrdquoEngineering Optimization vol 37 no 2 pp 167ndash183 2005
[35] K Deb and T Goel Controlled Elitist Non-Dominated SortingGenetic Algorithms for Better Convergence Springer BerlinGermany 2001
[36] Y J Lai T Y Liu and C L Hwang ldquoTopsis for MODMrdquoEuropean Journal of Operational Research vol 76 no 3 pp486ndash500 1994
Submit your manuscripts athttpwwwhindawicom
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical Problems in Engineering
Hindawi Publishing Corporationhttpwwwhindawicom
Differential EquationsInternational Journal of
Volume 2014
Applied MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Probability and StatisticsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical PhysicsAdvances in
Complex AnalysisJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
OptimizationJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
CombinatoricsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Operations ResearchAdvances in
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Function Spaces
Abstract and Applied AnalysisHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of Mathematics and Mathematical Sciences
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Algebra
Discrete Dynamics in Nature and Society
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Decision SciencesAdvances in
Discrete MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014 Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Stochastic AnalysisInternational Journal of
Journal of Applied Mathematics 11
Table 5 Notations for the model of the grinding and classificationprocess
Notation Description
119901119894
Particle percentage of the 119894th size fraction inthe ball mill overflow
119891119895
Particle percentage of the 119894th size fraction inthe Feed ore
119862 Rate of the first classifier recycle119864119886119894 The efficiency of the first spiral classifier
120591 Themean residence time1198751 The internal concentration in ball mill119872MF The solid flow of feed ore1198821 The water addition of the first classifier recycle
1198822 The classifier water addition
119887119894119895 The breakage distribution function119878119894 The breakage rate function119889119894 The particle with the 119894th size
120572119888119894
The particle percentage of the 119894th size in thefirst classifier overflow
119889119900 The unit size of the particle1198752 First-stage overflow120601119861 Ball filling rate
119860119888 119861119888
Parameters of the first-stage classifier overflowsize fraction distribution
119864119886119894 The efficiency of the first spiral classifier
11988950119888
The particle size fraction after correctionseparation
119898 The separation accuracy119896 The intermix index
1198641198861015840
min 1198891015840
50119888 1198981015840 1198961015840 The corresponding key parameters to theefficiency of the second spiral classifier
119886 120572 120583 ΛFour key parameters to control the breakagerate function
119889min 119889max Theminimum and maximum particle sizes119860MF 119861MF Parameters of feed ore size fraction distribution
119865(119894)The cumulative particle percentage less thanthe 119894th size fraction in feed ore
1205721198881015840
119894
The particle percentage of the 119894th size fractionin the second classifier overflow
1198641198861015840
119894 The efficiency of the second spiral classifier
42 Predictive Model of the Grinding and Classification Pro-cess Here we establish the mathematical predictive modelof each unit process in the bauxite grinding and classificationprocess The notations of the indexes decision variables andparameters are listed in Table 5 These notations will be usedfor the model of the grinding and classification process
421 Ball Mill Circuit Model Here 119901119894is the particle per-
centage of 119894th size fraction in the ball mill overflow 119891119895is the
particle percentage of 119894th size fraction in feed ore rate of thefirst classifier recycle 119862 is known and 119864119886
119894is the efficiency of
the first spiral classifier According to a technical report offield investigation and study we have that
119901119894(1 + 119862) =
119889119894119894119891119894+ sum119894minus1
119895=1119894gt1119889119894119895[119864119886119894119901119895(1 + 119862) + 119891
119895]
1 minus 119889119894119894119864119886119894
119889119894119895=
119890119895 119894 = 119895
119894minus1
sum
119896=119895
119888119894119896119888119895119896(119890119896minus 119890119894) 119894 gt 119895
119888119894119895=
minus
119895minus1
sum
119896=119894
119888119894119896119888119895119896 119894 lt 119895
1 119894 = 119895
1
119878119894minus 119878119895
119894minus1
sum
119896=119895
119878119896119887119894119896119888119896119895 119894 gt 119895
119890119895=
1
(1 + 05 sdot 120591119878119895) (1 + 025 sdot 120591119878
119895)2 120591 =
81198751
119872MF (12)
where 120591 is the mean residence time of minerals 1198751is the
internal concentration in ball mill and
1198751=
119872MF (1 + 119862)
119872MF (1 + 119862) +1198821 + 03 (1198821 +1198822) (13)
where119872MF (119905ℎ) is the solid flow of feed ore1198821is the water
addition of the first classifier recycle and1198822is the classifier
water addition 119887119894119895is the breakage distribution function 119878
119894
is the breakage rate function and it satisfied the followingequation
119878119894=
(119886(119889119894119889119900)120572
)
(1 + (119889119894120583)Λ
)
(14)
where 119889119894is the particle with the 119894th size 119889
119900it is a unit when
per millimeter is a unit 119889119900= 1 119889
119894= 119894 (mm) and 119886 120572 120583
and Λ are four key parameters to control the breakage ratefunction
In a concrete grinding and classification process the ballmill size is fixed and the speed of ball mill is constantThrough data acquisition and testing of grinding and clas-sification steady-state loop the regression model between 119886120572 120583 Λ and condition variables size fraction distributioncan be established The input variables are ball filling rate120601119861 solid flow of feed ore 119872MF water addition of the first
classifier recycle1198821 and parameters of feed ore size fraction
distribution 119860MF 119861MF The regression model is
[119886 120572 120583 Λ]119879
=[[
[
11990911
sdot sdot sdot 1199091119895
d
1199091198941
sdot sdot sdot 119909119894119895
]]
]
sdot [1198821 120601119861 119860MF 119861MF 119872MF]119879
(15)
The value of 119909119894119895(119894 = 1 2 3 4 119895 = 1 2 5) can be
obtained by the experimental data regression 119860MF 119861MF can
12 Journal of Applied Mathematics
be obtained from feed ore size fraction distribution and 119865(119894)is the cumulative particle percentage less than the 119894th sizefraction in feed ore and it is represented as follows
119865 (119894) = 1 minus exp (minus119860MF119889119861MF119894
) (16)
422 Spiral Classifier Model 120572119888119894is the particle percentage
of the 119894th size fraction in the first classifier overflow and 119901119894
is the particle percentage of the 119894th size fraction in ball milloverflow The spiral classifier model is as follows
120572119888119894=
119901119894times (1 minus 119864119886
119894)
sum119894=1
(119901119894times (1 minus 119864119886
119894))times 100 (17)
where 119864119886119894is the efficiency of the first spiral classifier and the
mechanism formula of 119864119886119894is shown as follows
119864119886119894= 1 minus exp [minus0693(
119889119894minus 119889min
11988950119888
minus 119889min)
119898
]
+ 119864119886min sdot [1 minus (119889119894minus 119889min119889max
)]
119896
(18)
where 119889119894is the particle with the 119894th size 119889min and 119889max
represent maximum and minimum particle sizes 11988950119888
isthe particle size fraction after correction separation 119898 isseparation accuracy and 119896 is intermix index
Through data acquisition and testing of grinding andclassification steady-state loop the regressionmodel betweenclassification parameters and condition variables size frac-tion distribution can be established The input variablesinclude the solid flow of feed ore 119872MF the classifier wateraddition 119882
2 and the parameters of ball mill overflow size
fraction distribution 119860MF 119861MF The regression model isshown as follows
[119864119886min 11988950119888
119898 119896]119879
=[[
[
11991011
sdot sdot sdot 1199101119895
d
1199101198941
sdot sdot sdot 119910119894119895
]]
]
sdot [119872MF 1198822119860MP 119861MP]
119879
(19)
where the value of 119910119894119895(119894 = 1 2 3 4 119895 = 1 2 3 4) can be
obtained by data regressionThe first-stage overflow 119875
2calculation formula is as
follows
1198752=
119872MF(119872MF +1198821 +1198822)
(20)
Similarly we can get the second spiral classifier model asfollows
1205721198881015840
119894=
120572119888119894times (1 minus 119864119886
1015840
119894)
sum119894(120572119888119894times (1 minus 1198641198861015840))
times 100 (21)
where 1205721198881015840119894is the particle percentage of the 119894th size fraction in
the second classifier overflow 120572119888119894is the particle percentage
of the 119894th size fraction in the first classifier overflow and
1198641198861015840
119894is the efficiency of the second spiral classifier The spiral
classifier model is as follows
1198641198861015840
119894= 1 minus exp[
[
minus0693(1198891015840
119894minus 1198891015840
min1198891015840
50119888minus 1198891015840
min)
1198981015840
]
]
+ 1198641198861015840
min sdot [1 minus (1198891015840
119894minus 1198891015840
min1198891015840max
)]
1198961015840
(22)
where 1198641198861015840min 1198891015840
50119888 1198981015840 and 119896
1015840 are key parameters to theefficiency of the second spiral classifier Through data acqui-sition and testing of grinding and classification steady-stateloop the regression model between classification parametersand condition variables size fraction distribution can beestablished The input variables include solid flow of feedore119872MF and parameters of the first-stage classifier overflowsize fraction distribution 119860
119888 119861119888 which are solved by similar
equation to (20) The regression model is shown as follows
[1198641198861015840
min 1198891015840
5011988811989810158401198961015840]119879
=
[[[
[
1199101015840
11sdot sdot sdot 1199101015840
1119895
d
1199101015840
1198941sdot sdot sdot 119910
1015840
119894119895
]]]
]
sdot [119872MF 119860119888119861119888]119879
(23)
where the value of1199101015840119894119895(119894 = 1 2 3 4 119895 = 1 2 3) can be obtained
by experimental data regression
43 Optimization Model of Grinding and Classification Pro-cess Two objective functions in the process are identifiedone is tomaximize output119891
1(119883) and the other is tomaximize
the small-size fractions (less than 0075mm fractions) in thesecond-stage overflow 119891
2(119883) It is also necessary to ensure
that the grinding productmeets all of other technical require-ments and the least disturbance in the following flotationcircuit As the constraints the feed load of the grinding circuit119872MF the steel ball filling rate 120601
119861 the first and the second
overflows 1198751and 119875
2 and the particle percentage of fine size
fraction in the first and the second classifier overflows 120572119888minus0075
and 1205721198881015840minus0075
should be within the user specified boundsThe operation variables are the solid flow of feed ore
119872MF water addition of the first classifier recycle 1198821 ball
filling rate 120601119861 and water addition of the second classifier119882
2
Based on all of the above grinding and classification processmultiobjective optimization model is as follows
max 119865 = max [1198911(119883) 119891
2(119883)]
1198911(119883) = 119872MF
1198912(119883) = 120572119888
1015840
minus0075
= 119891 (119872MF11988211198822 120601119861)
st 119872MFmin le 119872MF le 119872MFmax120601119861min le 120601119861 le 120601119861max1198751min le 1198751 le 1198751max1198752min le 1198752 le 1198752max120572119888minus0075
ge 120572119888min
1205721198881015840
minus0075ge 1205721198881015840
min
(24)
Journal of Applied Mathematics 13
Table 6 The optimization results calculated by the proposed algo-rithm
Number 1198911(119883) (th) 119891
2(119883) ()
1 92972 908142 91810 919003 92360 909234 90620 934605 91310 924946 90170 938007 89390 952008 89480 949329 91530 9240010 89298 9576311 89824 9454912 89800 9439513 89710 9461814 91170 9280015 91880 9184016 92190 9128117 89870 9409018 91000 9328519 91960 9152020 91124 93221
85
87
89
91
93
95
97
74000 79000 84000 94000
Optimization resultsThe raw data
f1(x)
f2(x)
()
89000
Figure 8 The comparison chart between optimization results andindustrial data
In (24) 119891(119872MF11988211198822 120601119861) implicates the model ofgrinding and classification represented by (12)ndash(23) Theproposed algorithm is applied to solve the problem and theoptimization results are shown in Table 6
With the practical process data from a grinding circuitof a mineral plant the simulation of this hybrid intelligentmethod adopted the same parameters on the variation infresh slurry feed velocity density particle size distributionand cyclone feed operating configurations
The comparison of production data and optimizationresults in Table 6 is shown in Figure 8 where ldquoXrdquo repre-sents the proposed algorithm optimization results and ldquoIrdquorepresents the original data collected from the field withoutoptimization of raw data According to the objectives thedata point closer to the upper right edge is more beneficialObviously the proposed optimization result is far betterthan the original data indicating the effectiveness of theoptimization approach
44 TOPSIS Method for Solution Selection The resolution ofamultiobjective optimization problemdoes not endwhen thePareto-optimal set is found In practical operational prob-lems a single solution must be selected TOPSIS [36] is auseful technique in dealing with multiattribute or multicrite-ria decision-making (MADMMCDM) problems in the realworld The standard TOPSIS method attempts to choosealternatives that simultaneously have the shortest distancefrom the positive-ideal solution and the farthest distancefrom the negative-ideal solution According to the TOPSISmethod the relative closeness coefficient is calculated andthe best solution in Table 6 is the solution number 10 as1198911(119883) = 89298 (119905ℎ) and 119891
2(119883) = 95763 () The
corresponding decision variables are 119872MF = 89298 (119905ℎ)120601119861= 32119882
1= 75127 (119905ℎ) and119882
2= 16296 (119905ℎ)
5 Conclusions
Promoted by the requirements of engineering optimizationin complex practical processes of grinding and classificationwe proposed a hybrid multiobjective differential evolutionalgorithm with a few beneficial features integrated Firstly anarchiving mechanism for infeasible solutions is establishedwith functional partitioned multi-population which aims todirect the population to approach or land in the feasibleregion from different directions during evolution Secondlywe propose an infeasible constraint violations function toselect infeasible population with better performance sothat they are allowed to be saved and to participate inthe subsequent evolution Thirdly a nondominating rankingstrategy is designed to improve the crowding-distance sortingand return uniform distribution of Pareto solutions Finallythe simplex method is inserted in the differential evolutionprocess to purposefully enrich the diversity without excessivecomputation cost The advantage of the proposed algorithmis the exemption from constructing penalty function andthe preservation of meaningful infeasible solutions directlySimulation results on benchmarks indicate that the proposedalgorithm can converge quickly and effectively to the truePareto frontier with better distribution
Based on the investigated information about grinding cir-cuit process we established a multiobjective optimal modelwith equations frommechanism knowledge parameters rec-ognized by data regression and constraints of technicalrequirements The nonlinear multiobjective optimizationmodel is too complicated to be solved by traditional gradient-based algorithmsThe proposed hybrid differential evolutionalgorithm is applied and tested to achieve a Pareto solution
14 Journal of Applied Mathematics
set It is proven to be valuable for operation decision makingin the industrial process and showed superiority to the oper-ation carried out in the production In fact many operatingparameters in complex processes are highly coupled andconflicting with each other The optimal operation of theentire production process is very difficult to obtain bymanualcalculation let alone the fluctuation situation of processconditions The application case indicates that the proposedmethod has good performance and is helpful to inspirefurther research on evolutionary methods for engineeringoptimization
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgments
This work is supported by the National Natural ScienceFoundation of China (61374156 61273187 and 61134006)the National Key Technology Research and DevelopmentProgram of the Ministry of Science and Technology ofChina (2012BAK09B04) the Science Fund for CreativeResearch Groups of the National Natural Science Foundationof China (61321003) the Fund for Doctor Station of theMinistry of Education (20110162130011 and 20100162120019)and theHunanProvince Science andTechnology PlanProject(2012CK4018)
References
[1] K Mitra and R Gopinath ldquoMultiobjective optimization ofan industrial grinding operation using elitist nondominatedsorting genetic algorithmrdquo Chemical Engineering Science vol59 no 2 pp 385ndash396 2004
[2] G Yu T Y Chai and X C Luo ldquoMultiobjective productionplanning optimization using hybrid evolutionary algorithmsfor mineral processingrdquo IEEE Transactions on EvolutionaryComputation vol 15 no 4 pp 487ndash514 2011
[3] T Y Ma W H Gui Y L Wang and C H Yang ldquoDynamicoptimization control for grinding and classification processrdquoControl andDecision vol 27 no 2 pp 286ndash290 2012 (Chinese)
[4] F Y Edgeworth Mathematical Physics An Essay on the Appli-cation of Mathematics to the Moral Sciences C Kegan Paul ampCompany London UK 1881
[5] V Pareto Cours drsquoEconomie Politique F Rouge LausanneSwitzerland 1896
[6] X F Chen W H Gui Y L Wang and L Cen ldquoMulti-stepoptimal control of complex process a genetic programmingstrategy and its applicationrdquo Engineering Applications of Artifi-cial Intelligence vol 17 no 5 pp 491ndash500 2004
[7] C A C Coello ldquoEvolutionary multi-objective optimizationa historical view of the fieldrdquo IEEE Computational IntelligenceMagazine vol 1 no 1 pp 28ndash36 2006
[8] K Deb A Pratap S Agarwal and T Meyarivan ldquoA fastand elitist multiobjective genetic algorithm NSGA-IIrdquo IEEETransactions on Evolutionary Computation vol 6 no 2 pp 182ndash197 2002
[9] C M Fonesca and P J Fleming ldquoGenetic algorithms formulti objective optimization formulation discussion and gen-eralizationrdquo in Proceedings of the 5th International Conferenceon Genetic Algorithms pp 415ndash423 Morgan Kaufmann SanMateo Calif USA 1993
[10] S Sengupta S Das M Nasir A V Vasilakos and W PedryczldquoEnergy-efficient differentiated coverage of dynamic objectsusing an improved evolutionary multi-objective optimizationalgorithm with fuzzy-dominancerdquo in Proceedings of IEEECongress on Evolutionary Computation pp 1ndash8 IEEE PressBrisbane Australia 2012
[11] K V Price R M Storn and J A Lampinen DifferentialEvolution a Practical Approach to Global Optimization SpringerBerlin Germany 2005
[12] W Y Gong and Z H Cai ldquoAn improved multiobjective dif-ferential evolution based on Pareto-adaptive 120576-dominance andorthogonal designrdquo European Journal of Operational Researchvol 198 no 2 pp 576ndash601 2009
[13] X H Zeng W K Wong and S Y Leung ldquoAn operatorallocation optimization model for balancing control of thehybrid assembly lines using Pareto utility discrete differentialevolution algorithmrdquo Computers and Operations Research vol39 no 5 pp 1145ndash1159 2012
[14] E Zitzler and L Thiele ldquoMultiobjective evolutionary algo-rithms a comparative case study and the strength Paretoapproachrdquo IEEETransactions onEvolutionaryComputation vol3 no 4 pp 257ndash271 1999
[15] A Al-Ani A Alsukker and R N Khushaba ldquoFeature subsetselection using differential evolution and a wheel based searchstrategyrdquo Swarm and Evolutionary Computation vol 9 pp 15ndash26 2013
[16] G Y Li and M G Liu ldquoThe summary of differential evolutionalgorithm and its improvementsrdquo in Proceedings of the 3rdInternational Conference on Advanced Computer Theory andEngineering (ICACTE rsquo10) pp V3153ndashV3156 Chengdu ChinaAugust 2010
[17] H A Abbass ldquoThe self-adaptive Pareto differential evolutionalgorithmrdquo in Proceedings of the IEEE Congress on EvolutionaryComputation pp 831ndash836 IEEE Press Honolulu Hawaii USAMay 2002
[18] N K Madavan ldquoMultiobjective optimization using a Paretodifferential evolution approachrdquo in Proceedings of the IEEECongress on Evolutionary Computation pp 1145ndash1150 IEEEPress Honolulu Hawaii USA May 2002
[19] F Xue A C Sanderson and R J Graves ldquoPareto-based multi-objective differential evolutionrdquo in Proceedings of the IEEECongress on Evolutionary Computation pp 862ndash869 IEEEPress Piscataway NJ USA December 2003
[20] T Robic and B Filipi ldquoDEMO differential evolution formultiobjective optimizationrdquo in Evolutionary Multi-CriterionOptimization pp 520ndash533 Springer Berlin Germany 2005
[21] V L Huang A K Qin P N Suganthan and M F TasgetirenldquoMulti-objective optimization based on self-adaptive differen-tial evolution algorithmrdquo in Proceedings of the IEEE Congresson Evolutionary Computation (CEC rsquo07) pp 3601ndash3608 IEEEPress Singapore September 2007
[22] V L Huang S Z Zhao R Mallipeddi and P N SuganthanldquoMulti-objective optimization using self-adaptive differentialevolution algorithmrdquo in Proceedings of the IEEE Congress onEvolutionary Computation (CEC rsquo09) pp 190ndash194 IEEE PressTrondheim Norway May 2009
Journal of Applied Mathematics 15
[23] J A Adeyemo and F A O Otieno ldquoMulti-objective differentialevolution algorithm for solving engineering problemsrdquo Journalof Applied Sciences vol 9 no 20 pp 3652ndash3661 2009
[24] Y Wang and Z Cai ldquoCombining multiobjective optimizationwith differential evolution to solve constrained optimizationproblemsrdquo IEEE Transactions on Evolutionary Computationvol 16 no 1 pp 117ndash134 2012
[25] A Homaifar C X Qi and S H Lai ldquoConstrained optimizationvia genetic algorithmsrdquo Simulation vol 62 no 4 pp 242ndash2531994
[26] J A Joines and C R Houck ldquoOn the use of non-stationarypenalty functions to solve nonlinear constrained optimizationproblems with GArsquosrdquo in Proceedings of the 1st IEEE Conferenceon Evolutionary Computation pp 579ndash584 Orlando Fla USAJune 1994
[27] F Jimenez and J L Verdegay ldquoEvolutionary techniques forconstrained optimization problemsrdquo in Proceedings of the 7thEuropean Congress on Intelligent Techniques and Soft Computing(EUFIT rsquo99) Aachen Germany 1999
[28] C Sun J Zeng and J Pan ldquoAn improved vector particleswarm optimization for constrained optimization problemsrdquoInformation Sciences vol 181 no 6 pp 1153ndash1163 2011
[29] Z Michalewicz and M Schoenauer ldquoEvolutionary algorithmsfor constrained parameter optimization problemsrdquo Evolution-ary Computation vol 4 no 1 pp 1ndash32 1996
[30] C A C Coello and A D Christiansen ldquoMoses a multiob-jective optimization tool for engineering designrdquo EngineeringOptimization vol 31 no 1ndash3 pp 337ndash368 1999
[31] Y G Woldesenbet G G Yen and B G Tessema ldquoConstrainthandling in multiobjective evolutionary optimizationrdquo IEEETransactions on Evolutionary Computation vol 13 no 3 pp514ndash525 2009
[32] J A Nelder and R Mead ldquoA simplex method for functionminimizationrdquoThe Computer Journal vol 7 no 4 pp 308ndash3131965
[33] D A van Veldhuizen and G B Lamont ldquoMultiobjective evolu-tionary algorithms analyzing the state-of-the-artrdquo EvolutionaryComputation vol 8 no 2 pp 125ndash147 2000
[34] R Farmani D A Savic andG AWalters ldquoEvolutionarymulti-objective optimization in water distribution network designrdquoEngineering Optimization vol 37 no 2 pp 167ndash183 2005
[35] K Deb and T Goel Controlled Elitist Non-Dominated SortingGenetic Algorithms for Better Convergence Springer BerlinGermany 2001
[36] Y J Lai T Y Liu and C L Hwang ldquoTopsis for MODMrdquoEuropean Journal of Operational Research vol 76 no 3 pp486ndash500 1994
Submit your manuscripts athttpwwwhindawicom
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical Problems in Engineering
Hindawi Publishing Corporationhttpwwwhindawicom
Differential EquationsInternational Journal of
Volume 2014
Applied MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Probability and StatisticsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical PhysicsAdvances in
Complex AnalysisJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
OptimizationJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
CombinatoricsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Operations ResearchAdvances in
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Function Spaces
Abstract and Applied AnalysisHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of Mathematics and Mathematical Sciences
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Algebra
Discrete Dynamics in Nature and Society
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Decision SciencesAdvances in
Discrete MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014 Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Stochastic AnalysisInternational Journal of
12 Journal of Applied Mathematics
be obtained from feed ore size fraction distribution and 119865(119894)is the cumulative particle percentage less than the 119894th sizefraction in feed ore and it is represented as follows
119865 (119894) = 1 minus exp (minus119860MF119889119861MF119894
) (16)
422 Spiral Classifier Model 120572119888119894is the particle percentage
of the 119894th size fraction in the first classifier overflow and 119901119894
is the particle percentage of the 119894th size fraction in ball milloverflow The spiral classifier model is as follows
120572119888119894=
119901119894times (1 minus 119864119886
119894)
sum119894=1
(119901119894times (1 minus 119864119886
119894))times 100 (17)
where 119864119886119894is the efficiency of the first spiral classifier and the
mechanism formula of 119864119886119894is shown as follows
119864119886119894= 1 minus exp [minus0693(
119889119894minus 119889min
11988950119888
minus 119889min)
119898
]
+ 119864119886min sdot [1 minus (119889119894minus 119889min119889max
)]
119896
(18)
where 119889119894is the particle with the 119894th size 119889min and 119889max
represent maximum and minimum particle sizes 11988950119888
isthe particle size fraction after correction separation 119898 isseparation accuracy and 119896 is intermix index
Through data acquisition and testing of grinding andclassification steady-state loop the regressionmodel betweenclassification parameters and condition variables size frac-tion distribution can be established The input variablesinclude the solid flow of feed ore 119872MF the classifier wateraddition 119882
2 and the parameters of ball mill overflow size
fraction distribution 119860MF 119861MF The regression model isshown as follows
[119864119886min 11988950119888
119898 119896]119879
=[[
[
11991011
sdot sdot sdot 1199101119895
d
1199101198941
sdot sdot sdot 119910119894119895
]]
]
sdot [119872MF 1198822119860MP 119861MP]
119879
(19)
where the value of 119910119894119895(119894 = 1 2 3 4 119895 = 1 2 3 4) can be
obtained by data regressionThe first-stage overflow 119875
2calculation formula is as
follows
1198752=
119872MF(119872MF +1198821 +1198822)
(20)
Similarly we can get the second spiral classifier model asfollows
1205721198881015840
119894=
120572119888119894times (1 minus 119864119886
1015840
119894)
sum119894(120572119888119894times (1 minus 1198641198861015840))
times 100 (21)
where 1205721198881015840119894is the particle percentage of the 119894th size fraction in
the second classifier overflow 120572119888119894is the particle percentage
of the 119894th size fraction in the first classifier overflow and
1198641198861015840
119894is the efficiency of the second spiral classifier The spiral
classifier model is as follows
1198641198861015840
119894= 1 minus exp[
[
minus0693(1198891015840
119894minus 1198891015840
min1198891015840
50119888minus 1198891015840
min)
1198981015840
]
]
+ 1198641198861015840
min sdot [1 minus (1198891015840
119894minus 1198891015840
min1198891015840max
)]
1198961015840
(22)
where 1198641198861015840min 1198891015840
50119888 1198981015840 and 119896
1015840 are key parameters to theefficiency of the second spiral classifier Through data acqui-sition and testing of grinding and classification steady-stateloop the regression model between classification parametersand condition variables size fraction distribution can beestablished The input variables include solid flow of feedore119872MF and parameters of the first-stage classifier overflowsize fraction distribution 119860
119888 119861119888 which are solved by similar
equation to (20) The regression model is shown as follows
[1198641198861015840
min 1198891015840
5011988811989810158401198961015840]119879
=
[[[
[
1199101015840
11sdot sdot sdot 1199101015840
1119895
d
1199101015840
1198941sdot sdot sdot 119910
1015840
119894119895
]]]
]
sdot [119872MF 119860119888119861119888]119879
(23)
where the value of1199101015840119894119895(119894 = 1 2 3 4 119895 = 1 2 3) can be obtained
by experimental data regression
43 Optimization Model of Grinding and Classification Pro-cess Two objective functions in the process are identifiedone is tomaximize output119891
1(119883) and the other is tomaximize
the small-size fractions (less than 0075mm fractions) in thesecond-stage overflow 119891
2(119883) It is also necessary to ensure
that the grinding productmeets all of other technical require-ments and the least disturbance in the following flotationcircuit As the constraints the feed load of the grinding circuit119872MF the steel ball filling rate 120601
119861 the first and the second
overflows 1198751and 119875
2 and the particle percentage of fine size
fraction in the first and the second classifier overflows 120572119888minus0075
and 1205721198881015840minus0075
should be within the user specified boundsThe operation variables are the solid flow of feed ore
119872MF water addition of the first classifier recycle 1198821 ball
filling rate 120601119861 and water addition of the second classifier119882
2
Based on all of the above grinding and classification processmultiobjective optimization model is as follows
max 119865 = max [1198911(119883) 119891
2(119883)]
1198911(119883) = 119872MF
1198912(119883) = 120572119888
1015840
minus0075
= 119891 (119872MF11988211198822 120601119861)
st 119872MFmin le 119872MF le 119872MFmax120601119861min le 120601119861 le 120601119861max1198751min le 1198751 le 1198751max1198752min le 1198752 le 1198752max120572119888minus0075
ge 120572119888min
1205721198881015840
minus0075ge 1205721198881015840
min
(24)
Journal of Applied Mathematics 13
Table 6 The optimization results calculated by the proposed algo-rithm
Number 1198911(119883) (th) 119891
2(119883) ()
1 92972 908142 91810 919003 92360 909234 90620 934605 91310 924946 90170 938007 89390 952008 89480 949329 91530 9240010 89298 9576311 89824 9454912 89800 9439513 89710 9461814 91170 9280015 91880 9184016 92190 9128117 89870 9409018 91000 9328519 91960 9152020 91124 93221
85
87
89
91
93
95
97
74000 79000 84000 94000
Optimization resultsThe raw data
f1(x)
f2(x)
()
89000
Figure 8 The comparison chart between optimization results andindustrial data
In (24) 119891(119872MF11988211198822 120601119861) implicates the model ofgrinding and classification represented by (12)ndash(23) Theproposed algorithm is applied to solve the problem and theoptimization results are shown in Table 6
With the practical process data from a grinding circuitof a mineral plant the simulation of this hybrid intelligentmethod adopted the same parameters on the variation infresh slurry feed velocity density particle size distributionand cyclone feed operating configurations
The comparison of production data and optimizationresults in Table 6 is shown in Figure 8 where ldquoXrdquo repre-sents the proposed algorithm optimization results and ldquoIrdquorepresents the original data collected from the field withoutoptimization of raw data According to the objectives thedata point closer to the upper right edge is more beneficialObviously the proposed optimization result is far betterthan the original data indicating the effectiveness of theoptimization approach
44 TOPSIS Method for Solution Selection The resolution ofamultiobjective optimization problemdoes not endwhen thePareto-optimal set is found In practical operational prob-lems a single solution must be selected TOPSIS [36] is auseful technique in dealing with multiattribute or multicrite-ria decision-making (MADMMCDM) problems in the realworld The standard TOPSIS method attempts to choosealternatives that simultaneously have the shortest distancefrom the positive-ideal solution and the farthest distancefrom the negative-ideal solution According to the TOPSISmethod the relative closeness coefficient is calculated andthe best solution in Table 6 is the solution number 10 as1198911(119883) = 89298 (119905ℎ) and 119891
2(119883) = 95763 () The
corresponding decision variables are 119872MF = 89298 (119905ℎ)120601119861= 32119882
1= 75127 (119905ℎ) and119882
2= 16296 (119905ℎ)
5 Conclusions
Promoted by the requirements of engineering optimizationin complex practical processes of grinding and classificationwe proposed a hybrid multiobjective differential evolutionalgorithm with a few beneficial features integrated Firstly anarchiving mechanism for infeasible solutions is establishedwith functional partitioned multi-population which aims todirect the population to approach or land in the feasibleregion from different directions during evolution Secondlywe propose an infeasible constraint violations function toselect infeasible population with better performance sothat they are allowed to be saved and to participate inthe subsequent evolution Thirdly a nondominating rankingstrategy is designed to improve the crowding-distance sortingand return uniform distribution of Pareto solutions Finallythe simplex method is inserted in the differential evolutionprocess to purposefully enrich the diversity without excessivecomputation cost The advantage of the proposed algorithmis the exemption from constructing penalty function andthe preservation of meaningful infeasible solutions directlySimulation results on benchmarks indicate that the proposedalgorithm can converge quickly and effectively to the truePareto frontier with better distribution
Based on the investigated information about grinding cir-cuit process we established a multiobjective optimal modelwith equations frommechanism knowledge parameters rec-ognized by data regression and constraints of technicalrequirements The nonlinear multiobjective optimizationmodel is too complicated to be solved by traditional gradient-based algorithmsThe proposed hybrid differential evolutionalgorithm is applied and tested to achieve a Pareto solution
14 Journal of Applied Mathematics
set It is proven to be valuable for operation decision makingin the industrial process and showed superiority to the oper-ation carried out in the production In fact many operatingparameters in complex processes are highly coupled andconflicting with each other The optimal operation of theentire production process is very difficult to obtain bymanualcalculation let alone the fluctuation situation of processconditions The application case indicates that the proposedmethod has good performance and is helpful to inspirefurther research on evolutionary methods for engineeringoptimization
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgments
This work is supported by the National Natural ScienceFoundation of China (61374156 61273187 and 61134006)the National Key Technology Research and DevelopmentProgram of the Ministry of Science and Technology ofChina (2012BAK09B04) the Science Fund for CreativeResearch Groups of the National Natural Science Foundationof China (61321003) the Fund for Doctor Station of theMinistry of Education (20110162130011 and 20100162120019)and theHunanProvince Science andTechnology PlanProject(2012CK4018)
References
[1] K Mitra and R Gopinath ldquoMultiobjective optimization ofan industrial grinding operation using elitist nondominatedsorting genetic algorithmrdquo Chemical Engineering Science vol59 no 2 pp 385ndash396 2004
[2] G Yu T Y Chai and X C Luo ldquoMultiobjective productionplanning optimization using hybrid evolutionary algorithmsfor mineral processingrdquo IEEE Transactions on EvolutionaryComputation vol 15 no 4 pp 487ndash514 2011
[3] T Y Ma W H Gui Y L Wang and C H Yang ldquoDynamicoptimization control for grinding and classification processrdquoControl andDecision vol 27 no 2 pp 286ndash290 2012 (Chinese)
[4] F Y Edgeworth Mathematical Physics An Essay on the Appli-cation of Mathematics to the Moral Sciences C Kegan Paul ampCompany London UK 1881
[5] V Pareto Cours drsquoEconomie Politique F Rouge LausanneSwitzerland 1896
[6] X F Chen W H Gui Y L Wang and L Cen ldquoMulti-stepoptimal control of complex process a genetic programmingstrategy and its applicationrdquo Engineering Applications of Artifi-cial Intelligence vol 17 no 5 pp 491ndash500 2004
[7] C A C Coello ldquoEvolutionary multi-objective optimizationa historical view of the fieldrdquo IEEE Computational IntelligenceMagazine vol 1 no 1 pp 28ndash36 2006
[8] K Deb A Pratap S Agarwal and T Meyarivan ldquoA fastand elitist multiobjective genetic algorithm NSGA-IIrdquo IEEETransactions on Evolutionary Computation vol 6 no 2 pp 182ndash197 2002
[9] C M Fonesca and P J Fleming ldquoGenetic algorithms formulti objective optimization formulation discussion and gen-eralizationrdquo in Proceedings of the 5th International Conferenceon Genetic Algorithms pp 415ndash423 Morgan Kaufmann SanMateo Calif USA 1993
[10] S Sengupta S Das M Nasir A V Vasilakos and W PedryczldquoEnergy-efficient differentiated coverage of dynamic objectsusing an improved evolutionary multi-objective optimizationalgorithm with fuzzy-dominancerdquo in Proceedings of IEEECongress on Evolutionary Computation pp 1ndash8 IEEE PressBrisbane Australia 2012
[11] K V Price R M Storn and J A Lampinen DifferentialEvolution a Practical Approach to Global Optimization SpringerBerlin Germany 2005
[12] W Y Gong and Z H Cai ldquoAn improved multiobjective dif-ferential evolution based on Pareto-adaptive 120576-dominance andorthogonal designrdquo European Journal of Operational Researchvol 198 no 2 pp 576ndash601 2009
[13] X H Zeng W K Wong and S Y Leung ldquoAn operatorallocation optimization model for balancing control of thehybrid assembly lines using Pareto utility discrete differentialevolution algorithmrdquo Computers and Operations Research vol39 no 5 pp 1145ndash1159 2012
[14] E Zitzler and L Thiele ldquoMultiobjective evolutionary algo-rithms a comparative case study and the strength Paretoapproachrdquo IEEETransactions onEvolutionaryComputation vol3 no 4 pp 257ndash271 1999
[15] A Al-Ani A Alsukker and R N Khushaba ldquoFeature subsetselection using differential evolution and a wheel based searchstrategyrdquo Swarm and Evolutionary Computation vol 9 pp 15ndash26 2013
[16] G Y Li and M G Liu ldquoThe summary of differential evolutionalgorithm and its improvementsrdquo in Proceedings of the 3rdInternational Conference on Advanced Computer Theory andEngineering (ICACTE rsquo10) pp V3153ndashV3156 Chengdu ChinaAugust 2010
[17] H A Abbass ldquoThe self-adaptive Pareto differential evolutionalgorithmrdquo in Proceedings of the IEEE Congress on EvolutionaryComputation pp 831ndash836 IEEE Press Honolulu Hawaii USAMay 2002
[18] N K Madavan ldquoMultiobjective optimization using a Paretodifferential evolution approachrdquo in Proceedings of the IEEECongress on Evolutionary Computation pp 1145ndash1150 IEEEPress Honolulu Hawaii USA May 2002
[19] F Xue A C Sanderson and R J Graves ldquoPareto-based multi-objective differential evolutionrdquo in Proceedings of the IEEECongress on Evolutionary Computation pp 862ndash869 IEEEPress Piscataway NJ USA December 2003
[20] T Robic and B Filipi ldquoDEMO differential evolution formultiobjective optimizationrdquo in Evolutionary Multi-CriterionOptimization pp 520ndash533 Springer Berlin Germany 2005
[21] V L Huang A K Qin P N Suganthan and M F TasgetirenldquoMulti-objective optimization based on self-adaptive differen-tial evolution algorithmrdquo in Proceedings of the IEEE Congresson Evolutionary Computation (CEC rsquo07) pp 3601ndash3608 IEEEPress Singapore September 2007
[22] V L Huang S Z Zhao R Mallipeddi and P N SuganthanldquoMulti-objective optimization using self-adaptive differentialevolution algorithmrdquo in Proceedings of the IEEE Congress onEvolutionary Computation (CEC rsquo09) pp 190ndash194 IEEE PressTrondheim Norway May 2009
Journal of Applied Mathematics 15
[23] J A Adeyemo and F A O Otieno ldquoMulti-objective differentialevolution algorithm for solving engineering problemsrdquo Journalof Applied Sciences vol 9 no 20 pp 3652ndash3661 2009
[24] Y Wang and Z Cai ldquoCombining multiobjective optimizationwith differential evolution to solve constrained optimizationproblemsrdquo IEEE Transactions on Evolutionary Computationvol 16 no 1 pp 117ndash134 2012
[25] A Homaifar C X Qi and S H Lai ldquoConstrained optimizationvia genetic algorithmsrdquo Simulation vol 62 no 4 pp 242ndash2531994
[26] J A Joines and C R Houck ldquoOn the use of non-stationarypenalty functions to solve nonlinear constrained optimizationproblems with GArsquosrdquo in Proceedings of the 1st IEEE Conferenceon Evolutionary Computation pp 579ndash584 Orlando Fla USAJune 1994
[27] F Jimenez and J L Verdegay ldquoEvolutionary techniques forconstrained optimization problemsrdquo in Proceedings of the 7thEuropean Congress on Intelligent Techniques and Soft Computing(EUFIT rsquo99) Aachen Germany 1999
[28] C Sun J Zeng and J Pan ldquoAn improved vector particleswarm optimization for constrained optimization problemsrdquoInformation Sciences vol 181 no 6 pp 1153ndash1163 2011
[29] Z Michalewicz and M Schoenauer ldquoEvolutionary algorithmsfor constrained parameter optimization problemsrdquo Evolution-ary Computation vol 4 no 1 pp 1ndash32 1996
[30] C A C Coello and A D Christiansen ldquoMoses a multiob-jective optimization tool for engineering designrdquo EngineeringOptimization vol 31 no 1ndash3 pp 337ndash368 1999
[31] Y G Woldesenbet G G Yen and B G Tessema ldquoConstrainthandling in multiobjective evolutionary optimizationrdquo IEEETransactions on Evolutionary Computation vol 13 no 3 pp514ndash525 2009
[32] J A Nelder and R Mead ldquoA simplex method for functionminimizationrdquoThe Computer Journal vol 7 no 4 pp 308ndash3131965
[33] D A van Veldhuizen and G B Lamont ldquoMultiobjective evolu-tionary algorithms analyzing the state-of-the-artrdquo EvolutionaryComputation vol 8 no 2 pp 125ndash147 2000
[34] R Farmani D A Savic andG AWalters ldquoEvolutionarymulti-objective optimization in water distribution network designrdquoEngineering Optimization vol 37 no 2 pp 167ndash183 2005
[35] K Deb and T Goel Controlled Elitist Non-Dominated SortingGenetic Algorithms for Better Convergence Springer BerlinGermany 2001
[36] Y J Lai T Y Liu and C L Hwang ldquoTopsis for MODMrdquoEuropean Journal of Operational Research vol 76 no 3 pp486ndash500 1994
Submit your manuscripts athttpwwwhindawicom
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical Problems in Engineering
Hindawi Publishing Corporationhttpwwwhindawicom
Differential EquationsInternational Journal of
Volume 2014
Applied MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Probability and StatisticsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical PhysicsAdvances in
Complex AnalysisJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
OptimizationJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
CombinatoricsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Operations ResearchAdvances in
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Function Spaces
Abstract and Applied AnalysisHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of Mathematics and Mathematical Sciences
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Algebra
Discrete Dynamics in Nature and Society
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Decision SciencesAdvances in
Discrete MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014 Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Stochastic AnalysisInternational Journal of
Journal of Applied Mathematics 13
Table 6 The optimization results calculated by the proposed algo-rithm
Number 1198911(119883) (th) 119891
2(119883) ()
1 92972 908142 91810 919003 92360 909234 90620 934605 91310 924946 90170 938007 89390 952008 89480 949329 91530 9240010 89298 9576311 89824 9454912 89800 9439513 89710 9461814 91170 9280015 91880 9184016 92190 9128117 89870 9409018 91000 9328519 91960 9152020 91124 93221
85
87
89
91
93
95
97
74000 79000 84000 94000
Optimization resultsThe raw data
f1(x)
f2(x)
()
89000
Figure 8 The comparison chart between optimization results andindustrial data
In (24) 119891(119872MF11988211198822 120601119861) implicates the model ofgrinding and classification represented by (12)ndash(23) Theproposed algorithm is applied to solve the problem and theoptimization results are shown in Table 6
With the practical process data from a grinding circuitof a mineral plant the simulation of this hybrid intelligentmethod adopted the same parameters on the variation infresh slurry feed velocity density particle size distributionand cyclone feed operating configurations
The comparison of production data and optimizationresults in Table 6 is shown in Figure 8 where ldquoXrdquo repre-sents the proposed algorithm optimization results and ldquoIrdquorepresents the original data collected from the field withoutoptimization of raw data According to the objectives thedata point closer to the upper right edge is more beneficialObviously the proposed optimization result is far betterthan the original data indicating the effectiveness of theoptimization approach
44 TOPSIS Method for Solution Selection The resolution ofamultiobjective optimization problemdoes not endwhen thePareto-optimal set is found In practical operational prob-lems a single solution must be selected TOPSIS [36] is auseful technique in dealing with multiattribute or multicrite-ria decision-making (MADMMCDM) problems in the realworld The standard TOPSIS method attempts to choosealternatives that simultaneously have the shortest distancefrom the positive-ideal solution and the farthest distancefrom the negative-ideal solution According to the TOPSISmethod the relative closeness coefficient is calculated andthe best solution in Table 6 is the solution number 10 as1198911(119883) = 89298 (119905ℎ) and 119891
2(119883) = 95763 () The
corresponding decision variables are 119872MF = 89298 (119905ℎ)120601119861= 32119882
1= 75127 (119905ℎ) and119882
2= 16296 (119905ℎ)
5 Conclusions
Promoted by the requirements of engineering optimizationin complex practical processes of grinding and classificationwe proposed a hybrid multiobjective differential evolutionalgorithm with a few beneficial features integrated Firstly anarchiving mechanism for infeasible solutions is establishedwith functional partitioned multi-population which aims todirect the population to approach or land in the feasibleregion from different directions during evolution Secondlywe propose an infeasible constraint violations function toselect infeasible population with better performance sothat they are allowed to be saved and to participate inthe subsequent evolution Thirdly a nondominating rankingstrategy is designed to improve the crowding-distance sortingand return uniform distribution of Pareto solutions Finallythe simplex method is inserted in the differential evolutionprocess to purposefully enrich the diversity without excessivecomputation cost The advantage of the proposed algorithmis the exemption from constructing penalty function andthe preservation of meaningful infeasible solutions directlySimulation results on benchmarks indicate that the proposedalgorithm can converge quickly and effectively to the truePareto frontier with better distribution
Based on the investigated information about grinding cir-cuit process we established a multiobjective optimal modelwith equations frommechanism knowledge parameters rec-ognized by data regression and constraints of technicalrequirements The nonlinear multiobjective optimizationmodel is too complicated to be solved by traditional gradient-based algorithmsThe proposed hybrid differential evolutionalgorithm is applied and tested to achieve a Pareto solution
14 Journal of Applied Mathematics
set It is proven to be valuable for operation decision makingin the industrial process and showed superiority to the oper-ation carried out in the production In fact many operatingparameters in complex processes are highly coupled andconflicting with each other The optimal operation of theentire production process is very difficult to obtain bymanualcalculation let alone the fluctuation situation of processconditions The application case indicates that the proposedmethod has good performance and is helpful to inspirefurther research on evolutionary methods for engineeringoptimization
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgments
This work is supported by the National Natural ScienceFoundation of China (61374156 61273187 and 61134006)the National Key Technology Research and DevelopmentProgram of the Ministry of Science and Technology ofChina (2012BAK09B04) the Science Fund for CreativeResearch Groups of the National Natural Science Foundationof China (61321003) the Fund for Doctor Station of theMinistry of Education (20110162130011 and 20100162120019)and theHunanProvince Science andTechnology PlanProject(2012CK4018)
References
[1] K Mitra and R Gopinath ldquoMultiobjective optimization ofan industrial grinding operation using elitist nondominatedsorting genetic algorithmrdquo Chemical Engineering Science vol59 no 2 pp 385ndash396 2004
[2] G Yu T Y Chai and X C Luo ldquoMultiobjective productionplanning optimization using hybrid evolutionary algorithmsfor mineral processingrdquo IEEE Transactions on EvolutionaryComputation vol 15 no 4 pp 487ndash514 2011
[3] T Y Ma W H Gui Y L Wang and C H Yang ldquoDynamicoptimization control for grinding and classification processrdquoControl andDecision vol 27 no 2 pp 286ndash290 2012 (Chinese)
[4] F Y Edgeworth Mathematical Physics An Essay on the Appli-cation of Mathematics to the Moral Sciences C Kegan Paul ampCompany London UK 1881
[5] V Pareto Cours drsquoEconomie Politique F Rouge LausanneSwitzerland 1896
[6] X F Chen W H Gui Y L Wang and L Cen ldquoMulti-stepoptimal control of complex process a genetic programmingstrategy and its applicationrdquo Engineering Applications of Artifi-cial Intelligence vol 17 no 5 pp 491ndash500 2004
[7] C A C Coello ldquoEvolutionary multi-objective optimizationa historical view of the fieldrdquo IEEE Computational IntelligenceMagazine vol 1 no 1 pp 28ndash36 2006
[8] K Deb A Pratap S Agarwal and T Meyarivan ldquoA fastand elitist multiobjective genetic algorithm NSGA-IIrdquo IEEETransactions on Evolutionary Computation vol 6 no 2 pp 182ndash197 2002
[9] C M Fonesca and P J Fleming ldquoGenetic algorithms formulti objective optimization formulation discussion and gen-eralizationrdquo in Proceedings of the 5th International Conferenceon Genetic Algorithms pp 415ndash423 Morgan Kaufmann SanMateo Calif USA 1993
[10] S Sengupta S Das M Nasir A V Vasilakos and W PedryczldquoEnergy-efficient differentiated coverage of dynamic objectsusing an improved evolutionary multi-objective optimizationalgorithm with fuzzy-dominancerdquo in Proceedings of IEEECongress on Evolutionary Computation pp 1ndash8 IEEE PressBrisbane Australia 2012
[11] K V Price R M Storn and J A Lampinen DifferentialEvolution a Practical Approach to Global Optimization SpringerBerlin Germany 2005
[12] W Y Gong and Z H Cai ldquoAn improved multiobjective dif-ferential evolution based on Pareto-adaptive 120576-dominance andorthogonal designrdquo European Journal of Operational Researchvol 198 no 2 pp 576ndash601 2009
[13] X H Zeng W K Wong and S Y Leung ldquoAn operatorallocation optimization model for balancing control of thehybrid assembly lines using Pareto utility discrete differentialevolution algorithmrdquo Computers and Operations Research vol39 no 5 pp 1145ndash1159 2012
[14] E Zitzler and L Thiele ldquoMultiobjective evolutionary algo-rithms a comparative case study and the strength Paretoapproachrdquo IEEETransactions onEvolutionaryComputation vol3 no 4 pp 257ndash271 1999
[15] A Al-Ani A Alsukker and R N Khushaba ldquoFeature subsetselection using differential evolution and a wheel based searchstrategyrdquo Swarm and Evolutionary Computation vol 9 pp 15ndash26 2013
[16] G Y Li and M G Liu ldquoThe summary of differential evolutionalgorithm and its improvementsrdquo in Proceedings of the 3rdInternational Conference on Advanced Computer Theory andEngineering (ICACTE rsquo10) pp V3153ndashV3156 Chengdu ChinaAugust 2010
[17] H A Abbass ldquoThe self-adaptive Pareto differential evolutionalgorithmrdquo in Proceedings of the IEEE Congress on EvolutionaryComputation pp 831ndash836 IEEE Press Honolulu Hawaii USAMay 2002
[18] N K Madavan ldquoMultiobjective optimization using a Paretodifferential evolution approachrdquo in Proceedings of the IEEECongress on Evolutionary Computation pp 1145ndash1150 IEEEPress Honolulu Hawaii USA May 2002
[19] F Xue A C Sanderson and R J Graves ldquoPareto-based multi-objective differential evolutionrdquo in Proceedings of the IEEECongress on Evolutionary Computation pp 862ndash869 IEEEPress Piscataway NJ USA December 2003
[20] T Robic and B Filipi ldquoDEMO differential evolution formultiobjective optimizationrdquo in Evolutionary Multi-CriterionOptimization pp 520ndash533 Springer Berlin Germany 2005
[21] V L Huang A K Qin P N Suganthan and M F TasgetirenldquoMulti-objective optimization based on self-adaptive differen-tial evolution algorithmrdquo in Proceedings of the IEEE Congresson Evolutionary Computation (CEC rsquo07) pp 3601ndash3608 IEEEPress Singapore September 2007
[22] V L Huang S Z Zhao R Mallipeddi and P N SuganthanldquoMulti-objective optimization using self-adaptive differentialevolution algorithmrdquo in Proceedings of the IEEE Congress onEvolutionary Computation (CEC rsquo09) pp 190ndash194 IEEE PressTrondheim Norway May 2009
Journal of Applied Mathematics 15
[23] J A Adeyemo and F A O Otieno ldquoMulti-objective differentialevolution algorithm for solving engineering problemsrdquo Journalof Applied Sciences vol 9 no 20 pp 3652ndash3661 2009
[24] Y Wang and Z Cai ldquoCombining multiobjective optimizationwith differential evolution to solve constrained optimizationproblemsrdquo IEEE Transactions on Evolutionary Computationvol 16 no 1 pp 117ndash134 2012
[25] A Homaifar C X Qi and S H Lai ldquoConstrained optimizationvia genetic algorithmsrdquo Simulation vol 62 no 4 pp 242ndash2531994
[26] J A Joines and C R Houck ldquoOn the use of non-stationarypenalty functions to solve nonlinear constrained optimizationproblems with GArsquosrdquo in Proceedings of the 1st IEEE Conferenceon Evolutionary Computation pp 579ndash584 Orlando Fla USAJune 1994
[27] F Jimenez and J L Verdegay ldquoEvolutionary techniques forconstrained optimization problemsrdquo in Proceedings of the 7thEuropean Congress on Intelligent Techniques and Soft Computing(EUFIT rsquo99) Aachen Germany 1999
[28] C Sun J Zeng and J Pan ldquoAn improved vector particleswarm optimization for constrained optimization problemsrdquoInformation Sciences vol 181 no 6 pp 1153ndash1163 2011
[29] Z Michalewicz and M Schoenauer ldquoEvolutionary algorithmsfor constrained parameter optimization problemsrdquo Evolution-ary Computation vol 4 no 1 pp 1ndash32 1996
[30] C A C Coello and A D Christiansen ldquoMoses a multiob-jective optimization tool for engineering designrdquo EngineeringOptimization vol 31 no 1ndash3 pp 337ndash368 1999
[31] Y G Woldesenbet G G Yen and B G Tessema ldquoConstrainthandling in multiobjective evolutionary optimizationrdquo IEEETransactions on Evolutionary Computation vol 13 no 3 pp514ndash525 2009
[32] J A Nelder and R Mead ldquoA simplex method for functionminimizationrdquoThe Computer Journal vol 7 no 4 pp 308ndash3131965
[33] D A van Veldhuizen and G B Lamont ldquoMultiobjective evolu-tionary algorithms analyzing the state-of-the-artrdquo EvolutionaryComputation vol 8 no 2 pp 125ndash147 2000
[34] R Farmani D A Savic andG AWalters ldquoEvolutionarymulti-objective optimization in water distribution network designrdquoEngineering Optimization vol 37 no 2 pp 167ndash183 2005
[35] K Deb and T Goel Controlled Elitist Non-Dominated SortingGenetic Algorithms for Better Convergence Springer BerlinGermany 2001
[36] Y J Lai T Y Liu and C L Hwang ldquoTopsis for MODMrdquoEuropean Journal of Operational Research vol 76 no 3 pp486ndash500 1994
Submit your manuscripts athttpwwwhindawicom
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical Problems in Engineering
Hindawi Publishing Corporationhttpwwwhindawicom
Differential EquationsInternational Journal of
Volume 2014
Applied MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Probability and StatisticsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical PhysicsAdvances in
Complex AnalysisJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
OptimizationJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
CombinatoricsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Operations ResearchAdvances in
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Function Spaces
Abstract and Applied AnalysisHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of Mathematics and Mathematical Sciences
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Algebra
Discrete Dynamics in Nature and Society
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Decision SciencesAdvances in
Discrete MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014 Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Stochastic AnalysisInternational Journal of
14 Journal of Applied Mathematics
set It is proven to be valuable for operation decision makingin the industrial process and showed superiority to the oper-ation carried out in the production In fact many operatingparameters in complex processes are highly coupled andconflicting with each other The optimal operation of theentire production process is very difficult to obtain bymanualcalculation let alone the fluctuation situation of processconditions The application case indicates that the proposedmethod has good performance and is helpful to inspirefurther research on evolutionary methods for engineeringoptimization
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgments
This work is supported by the National Natural ScienceFoundation of China (61374156 61273187 and 61134006)the National Key Technology Research and DevelopmentProgram of the Ministry of Science and Technology ofChina (2012BAK09B04) the Science Fund for CreativeResearch Groups of the National Natural Science Foundationof China (61321003) the Fund for Doctor Station of theMinistry of Education (20110162130011 and 20100162120019)and theHunanProvince Science andTechnology PlanProject(2012CK4018)
References
[1] K Mitra and R Gopinath ldquoMultiobjective optimization ofan industrial grinding operation using elitist nondominatedsorting genetic algorithmrdquo Chemical Engineering Science vol59 no 2 pp 385ndash396 2004
[2] G Yu T Y Chai and X C Luo ldquoMultiobjective productionplanning optimization using hybrid evolutionary algorithmsfor mineral processingrdquo IEEE Transactions on EvolutionaryComputation vol 15 no 4 pp 487ndash514 2011
[3] T Y Ma W H Gui Y L Wang and C H Yang ldquoDynamicoptimization control for grinding and classification processrdquoControl andDecision vol 27 no 2 pp 286ndash290 2012 (Chinese)
[4] F Y Edgeworth Mathematical Physics An Essay on the Appli-cation of Mathematics to the Moral Sciences C Kegan Paul ampCompany London UK 1881
[5] V Pareto Cours drsquoEconomie Politique F Rouge LausanneSwitzerland 1896
[6] X F Chen W H Gui Y L Wang and L Cen ldquoMulti-stepoptimal control of complex process a genetic programmingstrategy and its applicationrdquo Engineering Applications of Artifi-cial Intelligence vol 17 no 5 pp 491ndash500 2004
[7] C A C Coello ldquoEvolutionary multi-objective optimizationa historical view of the fieldrdquo IEEE Computational IntelligenceMagazine vol 1 no 1 pp 28ndash36 2006
[8] K Deb A Pratap S Agarwal and T Meyarivan ldquoA fastand elitist multiobjective genetic algorithm NSGA-IIrdquo IEEETransactions on Evolutionary Computation vol 6 no 2 pp 182ndash197 2002
[9] C M Fonesca and P J Fleming ldquoGenetic algorithms formulti objective optimization formulation discussion and gen-eralizationrdquo in Proceedings of the 5th International Conferenceon Genetic Algorithms pp 415ndash423 Morgan Kaufmann SanMateo Calif USA 1993
[10] S Sengupta S Das M Nasir A V Vasilakos and W PedryczldquoEnergy-efficient differentiated coverage of dynamic objectsusing an improved evolutionary multi-objective optimizationalgorithm with fuzzy-dominancerdquo in Proceedings of IEEECongress on Evolutionary Computation pp 1ndash8 IEEE PressBrisbane Australia 2012
[11] K V Price R M Storn and J A Lampinen DifferentialEvolution a Practical Approach to Global Optimization SpringerBerlin Germany 2005
[12] W Y Gong and Z H Cai ldquoAn improved multiobjective dif-ferential evolution based on Pareto-adaptive 120576-dominance andorthogonal designrdquo European Journal of Operational Researchvol 198 no 2 pp 576ndash601 2009
[13] X H Zeng W K Wong and S Y Leung ldquoAn operatorallocation optimization model for balancing control of thehybrid assembly lines using Pareto utility discrete differentialevolution algorithmrdquo Computers and Operations Research vol39 no 5 pp 1145ndash1159 2012
[14] E Zitzler and L Thiele ldquoMultiobjective evolutionary algo-rithms a comparative case study and the strength Paretoapproachrdquo IEEETransactions onEvolutionaryComputation vol3 no 4 pp 257ndash271 1999
[15] A Al-Ani A Alsukker and R N Khushaba ldquoFeature subsetselection using differential evolution and a wheel based searchstrategyrdquo Swarm and Evolutionary Computation vol 9 pp 15ndash26 2013
[16] G Y Li and M G Liu ldquoThe summary of differential evolutionalgorithm and its improvementsrdquo in Proceedings of the 3rdInternational Conference on Advanced Computer Theory andEngineering (ICACTE rsquo10) pp V3153ndashV3156 Chengdu ChinaAugust 2010
[17] H A Abbass ldquoThe self-adaptive Pareto differential evolutionalgorithmrdquo in Proceedings of the IEEE Congress on EvolutionaryComputation pp 831ndash836 IEEE Press Honolulu Hawaii USAMay 2002
[18] N K Madavan ldquoMultiobjective optimization using a Paretodifferential evolution approachrdquo in Proceedings of the IEEECongress on Evolutionary Computation pp 1145ndash1150 IEEEPress Honolulu Hawaii USA May 2002
[19] F Xue A C Sanderson and R J Graves ldquoPareto-based multi-objective differential evolutionrdquo in Proceedings of the IEEECongress on Evolutionary Computation pp 862ndash869 IEEEPress Piscataway NJ USA December 2003
[20] T Robic and B Filipi ldquoDEMO differential evolution formultiobjective optimizationrdquo in Evolutionary Multi-CriterionOptimization pp 520ndash533 Springer Berlin Germany 2005
[21] V L Huang A K Qin P N Suganthan and M F TasgetirenldquoMulti-objective optimization based on self-adaptive differen-tial evolution algorithmrdquo in Proceedings of the IEEE Congresson Evolutionary Computation (CEC rsquo07) pp 3601ndash3608 IEEEPress Singapore September 2007
[22] V L Huang S Z Zhao R Mallipeddi and P N SuganthanldquoMulti-objective optimization using self-adaptive differentialevolution algorithmrdquo in Proceedings of the IEEE Congress onEvolutionary Computation (CEC rsquo09) pp 190ndash194 IEEE PressTrondheim Norway May 2009
Journal of Applied Mathematics 15
[23] J A Adeyemo and F A O Otieno ldquoMulti-objective differentialevolution algorithm for solving engineering problemsrdquo Journalof Applied Sciences vol 9 no 20 pp 3652ndash3661 2009
[24] Y Wang and Z Cai ldquoCombining multiobjective optimizationwith differential evolution to solve constrained optimizationproblemsrdquo IEEE Transactions on Evolutionary Computationvol 16 no 1 pp 117ndash134 2012
[25] A Homaifar C X Qi and S H Lai ldquoConstrained optimizationvia genetic algorithmsrdquo Simulation vol 62 no 4 pp 242ndash2531994
[26] J A Joines and C R Houck ldquoOn the use of non-stationarypenalty functions to solve nonlinear constrained optimizationproblems with GArsquosrdquo in Proceedings of the 1st IEEE Conferenceon Evolutionary Computation pp 579ndash584 Orlando Fla USAJune 1994
[27] F Jimenez and J L Verdegay ldquoEvolutionary techniques forconstrained optimization problemsrdquo in Proceedings of the 7thEuropean Congress on Intelligent Techniques and Soft Computing(EUFIT rsquo99) Aachen Germany 1999
[28] C Sun J Zeng and J Pan ldquoAn improved vector particleswarm optimization for constrained optimization problemsrdquoInformation Sciences vol 181 no 6 pp 1153ndash1163 2011
[29] Z Michalewicz and M Schoenauer ldquoEvolutionary algorithmsfor constrained parameter optimization problemsrdquo Evolution-ary Computation vol 4 no 1 pp 1ndash32 1996
[30] C A C Coello and A D Christiansen ldquoMoses a multiob-jective optimization tool for engineering designrdquo EngineeringOptimization vol 31 no 1ndash3 pp 337ndash368 1999
[31] Y G Woldesenbet G G Yen and B G Tessema ldquoConstrainthandling in multiobjective evolutionary optimizationrdquo IEEETransactions on Evolutionary Computation vol 13 no 3 pp514ndash525 2009
[32] J A Nelder and R Mead ldquoA simplex method for functionminimizationrdquoThe Computer Journal vol 7 no 4 pp 308ndash3131965
[33] D A van Veldhuizen and G B Lamont ldquoMultiobjective evolu-tionary algorithms analyzing the state-of-the-artrdquo EvolutionaryComputation vol 8 no 2 pp 125ndash147 2000
[34] R Farmani D A Savic andG AWalters ldquoEvolutionarymulti-objective optimization in water distribution network designrdquoEngineering Optimization vol 37 no 2 pp 167ndash183 2005
[35] K Deb and T Goel Controlled Elitist Non-Dominated SortingGenetic Algorithms for Better Convergence Springer BerlinGermany 2001
[36] Y J Lai T Y Liu and C L Hwang ldquoTopsis for MODMrdquoEuropean Journal of Operational Research vol 76 no 3 pp486ndash500 1994
Submit your manuscripts athttpwwwhindawicom
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical Problems in Engineering
Hindawi Publishing Corporationhttpwwwhindawicom
Differential EquationsInternational Journal of
Volume 2014
Applied MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Probability and StatisticsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical PhysicsAdvances in
Complex AnalysisJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
OptimizationJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
CombinatoricsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Operations ResearchAdvances in
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Function Spaces
Abstract and Applied AnalysisHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of Mathematics and Mathematical Sciences
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Algebra
Discrete Dynamics in Nature and Society
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Decision SciencesAdvances in
Discrete MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014 Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Stochastic AnalysisInternational Journal of
Journal of Applied Mathematics 15
[23] J A Adeyemo and F A O Otieno ldquoMulti-objective differentialevolution algorithm for solving engineering problemsrdquo Journalof Applied Sciences vol 9 no 20 pp 3652ndash3661 2009
[24] Y Wang and Z Cai ldquoCombining multiobjective optimizationwith differential evolution to solve constrained optimizationproblemsrdquo IEEE Transactions on Evolutionary Computationvol 16 no 1 pp 117ndash134 2012
[25] A Homaifar C X Qi and S H Lai ldquoConstrained optimizationvia genetic algorithmsrdquo Simulation vol 62 no 4 pp 242ndash2531994
[26] J A Joines and C R Houck ldquoOn the use of non-stationarypenalty functions to solve nonlinear constrained optimizationproblems with GArsquosrdquo in Proceedings of the 1st IEEE Conferenceon Evolutionary Computation pp 579ndash584 Orlando Fla USAJune 1994
[27] F Jimenez and J L Verdegay ldquoEvolutionary techniques forconstrained optimization problemsrdquo in Proceedings of the 7thEuropean Congress on Intelligent Techniques and Soft Computing(EUFIT rsquo99) Aachen Germany 1999
[28] C Sun J Zeng and J Pan ldquoAn improved vector particleswarm optimization for constrained optimization problemsrdquoInformation Sciences vol 181 no 6 pp 1153ndash1163 2011
[29] Z Michalewicz and M Schoenauer ldquoEvolutionary algorithmsfor constrained parameter optimization problemsrdquo Evolution-ary Computation vol 4 no 1 pp 1ndash32 1996
[30] C A C Coello and A D Christiansen ldquoMoses a multiob-jective optimization tool for engineering designrdquo EngineeringOptimization vol 31 no 1ndash3 pp 337ndash368 1999
[31] Y G Woldesenbet G G Yen and B G Tessema ldquoConstrainthandling in multiobjective evolutionary optimizationrdquo IEEETransactions on Evolutionary Computation vol 13 no 3 pp514ndash525 2009
[32] J A Nelder and R Mead ldquoA simplex method for functionminimizationrdquoThe Computer Journal vol 7 no 4 pp 308ndash3131965
[33] D A van Veldhuizen and G B Lamont ldquoMultiobjective evolu-tionary algorithms analyzing the state-of-the-artrdquo EvolutionaryComputation vol 8 no 2 pp 125ndash147 2000
[34] R Farmani D A Savic andG AWalters ldquoEvolutionarymulti-objective optimization in water distribution network designrdquoEngineering Optimization vol 37 no 2 pp 167ndash183 2005
[35] K Deb and T Goel Controlled Elitist Non-Dominated SortingGenetic Algorithms for Better Convergence Springer BerlinGermany 2001
[36] Y J Lai T Y Liu and C L Hwang ldquoTopsis for MODMrdquoEuropean Journal of Operational Research vol 76 no 3 pp486ndash500 1994
Submit your manuscripts athttpwwwhindawicom
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical Problems in Engineering
Hindawi Publishing Corporationhttpwwwhindawicom
Differential EquationsInternational Journal of
Volume 2014
Applied MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Probability and StatisticsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical PhysicsAdvances in
Complex AnalysisJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
OptimizationJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
CombinatoricsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Operations ResearchAdvances in
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Function Spaces
Abstract and Applied AnalysisHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of Mathematics and Mathematical Sciences
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Algebra
Discrete Dynamics in Nature and Society
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Decision SciencesAdvances in
Discrete MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014 Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Stochastic AnalysisInternational Journal of
Submit your manuscripts athttpwwwhindawicom
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical Problems in Engineering
Hindawi Publishing Corporationhttpwwwhindawicom
Differential EquationsInternational Journal of
Volume 2014
Applied MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Probability and StatisticsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical PhysicsAdvances in
Complex AnalysisJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
OptimizationJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
CombinatoricsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Operations ResearchAdvances in
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Function Spaces
Abstract and Applied AnalysisHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of Mathematics and Mathematical Sciences
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Algebra
Discrete Dynamics in Nature and Society
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Decision SciencesAdvances in
Discrete MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014 Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Stochastic AnalysisInternational Journal of