a brkga-based matheuristic for the maximum quasi …celso/artigos/brkga_exactqclique.pdf · imum...
TRANSCRIPT
-
RAIRO Operations ResearchWill be set by the publisher
A BRKGA-BASED MATHEURISTIC FOR THE MAXIMUMQUASI-CLIQUE PROBLEM WITH AN EXACT LOCAL
SEARCH STRATEGY
Bruno Q. Pinto1, Celso C. Ribeiro2, José A. Riveaux3 andIsabel Rosseti4
Abstract. Given a graph G = (V,E) and a threshold γ ∈ (0, 1], themaximum cardinality quasi-clique problem consists in finding a maxi-mum cardinality subset C∗ of the vertices in V such that the densityof the graph induced in G by C∗ is greater than or equal to the thresh-old γ. This problem has a number of applications in data mining, e.g.in social networks or phone call graphs. We propose a matheuristicfor solving the maximum cardinality quasi-clique problem, based onthe hybridization of a biased random-key genetic algorithm (BRKGA)with an exact local search strategy. The newly proposed approach iscompared with a pure biased random-key genetic algorithm, which wasthe best heuristic in the literature at the time of writing. Compu-tational results show that the hybrid BRKGA outperforms the pureBRKGA.
Keywords: maximum cardinality quasi-clique problem, maximumclique problem, biased random-key genetic algorithm, metaheuristics,matheuristic, graph density.
June 7, 2019.1 Instituto Federal de Educação, Ciência e Tecnologia do Triângulo Mineiro, Uberlândia, MG38411-104, Brazil. e-mail: [email protected] Institute of Computing, Universidade Federal Fluminense, Niterói, RJ 24210-346, Brazil.e-mail: [email protected] Institute of Computing, Universidade Federal Fluminense, Niterói, RJ 24210-346, Brazil.e-mail: [email protected] Institute of Computing, Universidade Federal Fluminense, Niterói, RJ 24210-346, Brazil.e-mail: [email protected]
c© EDP Sciences 2019
-
2 TITLE WILL BE SET BY THE PUBLISHER
1. Introduction
Let G = (V,E) be a graph defined by a vertex set V and an edge set E ⊆ V ×V .G is a complete graph if there is an edge in E connecting every two different verticesin V . A graph G′ = (V ′, E′) is a subgraph of G if V ′ ⊆ V and E′ ⊆ E, whichis denoted by G′ ⊆ G. The graph G(V ′) induced in G by V ′ ⊆ V is that withvertex set V ′ and edge set formed by all edges of E with both ends in V ′. For anyV ′ ⊆ V , the subset E(V ′) ⊆ E is formed by all edges of E with both ends in V ′.In other words, E(V ′) is the edge set of the graph induced in G by V ′.
The density of graph G is given by dens(G) = |E|/(|V | × (|V | − 1)/2). For anyv ∈ V , the degree degG(v) denotes the number of vertices in G that are adjacentto v. In addition, for any v ∈ V and any V ′ ⊆ V , we denote by degG(v, V ′) thenumber of vertices of V ′ that are adjacent to v in G.
A subset C ⊆ V is a clique ofG if the graphG(C) induced inG by C is complete.Given a graph G = (V,E), the maximum clique problem (MCP) consists in findinga maximum cardinality clique of G. It was proved to be NP-hard by [15].
Given a graph G = (V,E) and a threshold γ ∈ (0, 1], a γ-clique is any subsetC ⊆ V such that the density of the subgraph G(C) is greater than or equal toγ. A γ-clique C is maximal if there is no other γ-clique C ′ that strictly containsC. The maximum quasi-clique problem (MQCP) amounts to finding a maximumcardinality subset C∗ of the vertices in V such that the density of the graphinduced in G by C∗ is greater than or equal to the threshold γ. This problem wasproved to be NP-hard in [20]. The problem has many applications and relatedclustering approaches include classifying molecular sequences in genome projectsby using a linkage graph of their pairwise similarities [8] and the analysis of massivetelecommunication data sets obtained from social networks or phone call graphs [1],as well as various other data mining and graph mining applications.
A few heuristics for MQCP exist in the literature, based on well known ap-proaches such as greedy randomized algorithms and their iterated extensions [18],stochastic local search [8], and GRASP [1]. Pinto et al. [22] proposed a biasedrandom-key genetic algorithm for finding approximate solutions to the maximumcardinality quasi-clique problem, using two different decoders. They showed thatthe decoder based on an optimized iterated greedy constructive heuristic led tothe best numerical results. They also showed that the use of a restart strat-egy significantly contributed to improve the robustness and the efficiency of thealgorithm. The resulting BRKGA-IG∗ heuristic strategy achieved the best perfor-mance and outperformed the restarted optimized iterated greedy (RIG∗) construc-tion/destruction heuristic of Oliveira et al. [18]. BRKGA-IG∗ was also comparedwith the exact algorithms AlgF3 and AlgF4 of Veremyev et al. [29] used as heuris-tics with time limits on their running times. BRKGA-IG∗ applied to sparse graphsalso outperformed the mixed integer programming approaches, finding target so-lution values in much smaller running times.
Ribeiro and Riveaux [24] proposed an exact enumeration algorithm to solve themaximum quasi-clique problem, based on a quasi-hereditary property. They alsoproposed a new upper bound that is used for pruning the search tree. Numerical
-
TITLE WILL BE SET BY THE PUBLISHER 3
results showed that their approach is competitive with the best integer program-ming formulations in [20, 29] solved by CPLEX and with the branch-and-boundalgorithm proposed by [19], in terms of both solution quality and running time.
We show that the exact enumeration algorithm QClique proposed by Ribeiroand Riveaux [24] can be hybridized with the biased random-key genetic algorithmBRKGA-IG∗ developed by Pinto et al. [22] as a local search strategy to improvethe quality of the solutions created by the decoder. This paper is organized asfollows. Section 2 presents the problem formulation. Section 3 introduces biasedrandom-key genetic algorithms and describes their customization to the maximumquasi-clique problem. Section 4 describes in detail the decoder DECODER-IG∗previously used used in the implementation of the biased random-key genetic al-gorithm for the maximum quasi-clique problem. The new decoder DECODER-LSQClique using an exact local search strategy is presented in Section 5, givingrise to a matheuristic for solving the maximum cardinality quasi-clique problembased on a biased random-key genetic algorithm. The exact local search algorithmLSQClique used by the new decoder is described in detail in Section 6. Numer-ical results are reported in Section 7. Concluding remarks are drawn in the lastsection.
2. Problem formulation and related work
The maximum quasi-clique problem can be formulated by associating a binaryvariable xi to each vertex of the graph [20]:
xi =
{1, if vertex vi ∈ V belongs to the solution,0, otherwise.
This formulation also considers a variable yij = xi · xj associated to each pairof vertices i, j ∈ V , with i < j, which is linearized as follows:
max∑i∈V
xi (1)
subject to: ∑(i,j)∈E:i
-
4 TITLE WILL BE SET BY THE PUBLISHER
ensures that the density of the solution is greater than or equal to γ. Constraints(3) and (4) ensure that any edge may contribute to the density of a solution onlyif both of its ends are chosen to belong to this solution. Constraints (5) ensurethat any existing edge (i, j) ∈ E will contribute to the solution if both of itsends are chosen. Constraints (6) and (7) impose the binary and non-negativityrequirements on the problem variables, respectively.
Veremyev et al. [29] reported and compared four mixed integer programmingformulations for the maximum quasi-clique problem in sparse graphs. Two algo-rithms based on the best formulations led to better results than the mixed integerprogramming formulation proposed in [20], with all mixed integer programs solvedusing FICO Xpress-Optimizer [10] with the time limit of 3600 seconds. Ribeiroand Riveaux [24] developed an exact algorithm based on a quasi-hereditary prop-erty and proposed a new upper bound that is used for pruning the search tree.Numerical results showed that their approach is competitive and outperforms thebest integer programming approaches in the literature. The new upper bound isconsistently tighter than previously existing bounds.
3. Biased random-key genetic algorithms for maximumquasi-clique
Genetic algorithms with random keys, also called random-key genetic algo-rithms (RKGA), were first introduced by Bean [4] for combinatorial optimizationproblems whose solutions may be represented by permutation vectors. Solutionsare represented as vectors of randomly generated real numbers called keys. Adeterministic algorithm, called a decoder, takes as input a solution vector andassociates with it a feasible solution of the combinatorial optimization problem,for which an objective value or fitness can be computed. Two parents are selectedat random from the entire population to implement the crossover operation in theimplementation of an RKGA. Parents are allowed to be selected for mating morethan once in the same generation.
A biased random-key genetic algorithm (BRKGA) differs from an RKGA inthe way parents are selected for crossover, see Gonçalves and Resende [11] andResende and Ribeiro [23] for reviews. In a BRKGA, each element is generatedcombining one element selected at random from the elite solutions in the currentpopulation, while the other is a non-elite solution. The selection is said to bebiased because one parent is always an elite solution and has a higher probabilityof passing its genes to the new generation.
In the following, we summarize two variants of a biased random-key geneticalgorithm for MQCP, each of them using a different decoder. Both of them evolvea population of chromosomes that consists of vectors of real numbers. Each chro-mosome is represented by a vector of |V | components, in which each key is a realnumber in the range [0, 1) associated with one of the vertices of the graph G. Eachchromosome is decoded by an algorithm that receives the vector of keys and buildsa feasible solution for MQCP, i.e., the decoder returns a γ-clique as its output.
-
TITLE WILL BE SET BY THE PUBLISHER 5
The two decoders DECODER-HCB and DECODER-IG∗ are described in the nextsection and have been originally presented in [22].
The parametric uniform crossover scheme proposed by [27] is used to combinetwo parent solutions and to produce an offspring. In this scheme, the offspringinherits each of its keys from the best fit of the two parents with a higher proba-bility. The biased random-key genetic algorithm developed in this work does notmake use of the standard mutation operator, where parts of the chromosomes arechanged with small probability. Instead, the concept of mutants is used: mutantsolutions are introduced in the population in each generation, randomly generatedin the same way as in the initial population. Mutants play the same role of themutation operator in traditional genetic algorithms, diversifying the search andhelping the procedure to escape from locally optimal solutions ([6, 7, 17]).
The |V | keys in the chromosome are randomly generated in the initial popula-tion. At each generation, the population is partitioned into two sets: TOP andREST . The size of the population is |TOP |+ |REST |. Subset TOP contains thebest solutions in the population. Subset REST is formed by two disjoint subsets:MID and BOT , with subset BOT being formed by the worst elements in the cur-rent population. As illustrated in Figure 1, the chromosomes in TOP are simplycopied to the population of the next generation. The elements in BOT are replacedby newly created mutants that are placed in the new set BOT . The remaining ele-ments of the new population are obtained by crossover, with one parent randomlychosen from TOP and the other from REST . This distinguishes a biased random-key genetic algorithm from the random-key genetic algorithm of [4], where bothparents are selected at random from the entire population. Since a parent solutioncan be chosen for crossover more than once in any given generation, elite solutionshave a higher probability of passing their random keys to the next generation. Inthis way, |MID | = |REST | − |BOT | offspring solutions are created.
The implementations of the biased random-key genetic algorithms for the max-imum quasi-clique problem make use of the C++ library brkgaAPI developedby [28], which is a framework for the development of biased random-key geneticalgorithms. It can also be used in parallel architectures running OpenMP.
The instantiation of the framework shown in Figure 2 to some specific opti-mization problem requires exclusively the development of a class implementingthe decoder for this problem. This is the only problem-dependent part of the tool.
According to [12], the BRKGA framework requires the following parameters:(a) the population size (p = |TOP | + |REST |); (b) the fraction pe of the popu-lation corresponding to the elite set TOP ; (c) the fraction pm of the populationcorresponding to the mutant set BOT ; (d) the probability rhoe that the offspringinherits each of its keys from the best fit of the two parents; and (e) the num-ber k of generations without improvement in the best solution until a restart isperformed.
In the remainder of this work, we consider and compare two variants of a bi-ased random-key genetic algorithm for solving MQCP, each of them based on adifferent decoder. Decoder DECODER-IG∗ was originally proposed by [22] andwill be summarized in the next section. The new decoder DECODER-LSQClique
-
6 TITLE WILL BE SET BY THE PUBLISHER
TOP TOP
REST
BOT
XCrossover
Copy best solutions
Select one parentfrom TOP
Select otherparent from REST
Randomly generatedsolutions
most fit
least fit
MID
Figure 1. Population evolution between consecutive generationsof a BRKGA.
begingenerate P vectors
of random keys
stopping rulesatisfied?
decode each vectorof random keys
sort solutions bytheir fitness
classify solutions as
elite or non-elite
copy elite solutionsto next population
generate mutants innext population
combine selected parentsand add offspringto next population
update best
result
restart rulesatisfied?
yes
no
endyes
no
Figure 2. BRKGA framework.
proposed in this work is based on the exact enumeration algorithm proposed by[24] and will be presented in Section 5.
4. Decoder DECODER-IG∗
We first review decoder DECODER-HCB, as presented in [22]. Each solutionis associated with a set of |V | random keys. The decoder receives as parameters
-
TITLE WILL BE SET BY THE PUBLISHER 7
the random keys rj ∈ [0, 1), j = 1, . . . , |V |. Each random key is a real numberin the range [0, 1) and corresponds to a vertex of the graph. Each chromosomerepresented by a set of random keys is decoded by an algorithm that receives thekeys and builds a feasible solution to MQCP. In other words, the decoder returnsa γ-clique associated with the set of random keys. Its pseudo-code is describedin Algorithm 1. The roles of parameters minsize and α are the same explainedin [22].
It may be used in two situations. First, to build a solution from scratch. Second,to complete (i.e., to reconstruct) a partially destroyed solution. For the secondcase, the decoder receives as an additional parameter a partial solution formed bya non-empty list of vertices.
Algorithm 1 DECODER-HCB(G, γ, α,minsize, S, r)
1: CL← V \ S2: if S = ∅ then3: RCL← {v ∈ CL : |{v′ ∈ CL : degG(v′) ≥ degG(v)}| ≤ max{li , α · |CL|}}4: x← argmin{rj : j ∈ RCL}5: S ← {x}6: end if7: while CL 6= ∅ do8: CL← ∅9: for all v ∈ V \ S do
10: if 2·(|E(S)|+degG(v,S))|S|(|S|+1) ≥ γ then11: CL← CL ∪ {v}12: end if13: end for14: if CL 6= ∅ then15: for all v ∈ CL do16: dif (v)← degG(v,CL) + |CL| · (degG(v, S)− γ · (|S|+ 1))17: end for18: RCL ← {v ∈ CL : |{v′ ∈ CL : dif (v′) ≥ dif (v)}| ≤ max{minsize, α ·|CL|}}
19: x← argmin{rj : j ∈ RCL}20: S ← S ∪ {x}21: end if22: end while
The second decoder DECODER-IG∗ is an extension of DECODER-HCB pre-sented above. It is based on the constructive heuristic HCB and on the optimizediterated, but decodes a population formed by longer vector of 2 · |V | random keysRj ∈ [0, 1), j = 1, . . . , |V |, |V | + 1, . . . , 2 · |V | each. The first |V | positions of eachvector of random keys are used in the construction of the initial solution and inthe reconstruction phase, while the last |V | positions are used in the destructionphase. The roles of parameters minsize, α, δ, and β are explained in [22].
-
8 TITLE WILL BE SET BY THE PUBLISHER
The pseudo-code of Algorithm 2 starts by creating an initial solution S′ in line 1,using the decoder DECODER-HCB and the first random keys Rj , j = 1, . . . , |V |.The loop in lines 2 to 10 repeats the partial destruction (vertex eliminations)followed by the reconstruction (vertex insertions) of the current solution, until nofurther improvements can be obtained. The current solution S′ is copied to S inline 2. The current solution S′ is copied to S in line 3. The loop in lines 4 to 8removes one by one the δ · |S′| vertices that should be eliminated from the currentsolution. A restricted candidate RCL of size max{minsize, β · |CL|} is createdin line 5, containing the vertices with the smallest degrees in G(S′). The vertexwith the smallest random key R|V |+j , j ∈ RCL, is selected from the restrictedcandidate list in line 6 and eliminated from the current solution in line 7. Thereconstruction phase is performed in line 9, where the current partial solution S′is rebuilt by decoder DECODER-HCB, once again using the first random keysRj ,∈ RCL. The loop is interrupted in line 10 when the new solution S′ obtainedby destruction-reconstruction does not improve the incumbent S or the graphG(S′) is not connected; otherwise a new iteration resumes.
Algorithm 2 DECODER-IG∗(G, γ, α, δ, β,minsize, S,R)
1: S′ ← DECODER-HCB(G, γ, α,minsize, ∅, Rj : j = 1, . . . , |V |)2: repeat3: S ← S′4: for k = 1 to δ · |S′| do5: RCL ← {v ∈ S′ : |{v′ ∈ S′ : degG(v′, S′) ≤ degG(v, S′)}| ≤
max{minsize, β · |S ′|}}6: x← argmin{R|V |+j : j ∈ RCL}7: S′ ← S′ \ {x}8: end for9: S′ ← DECODER-HCB(G, γ, α,minsize, S′, Rj : j = 1, . . . , |V |)
10: until |S′| ≤ |S| or graph G(S′) is not connected
5. New decoder based on exact local search
The principle of the new decoder proposed in this section consists in replacingthe reconstruction phase of decoder DECODER-IG∗ (i.e., line 9 of Algorithm 2) byan exact algorithm that finds the maximum quasi-clique that can be reconstructedby DECODER-HCB starting from the remaining set of vertices S′ with α = 1. Aslightly modified version of the QClique algorithm of [24] is used with this goal.
Algorithm 3 describes the pseudo-code of decoder DECODER-LSQClique. Itrequires the same parameters as DECODER-IG∗, with the exception of an ex-tended vector with 2 · |V | + 1 random keys R+j , j = 1, . . . , 2 · |V |, 2 · |V | + 1 anda frequency parameter ρ: the exact search algorithm will be applied wheneverR+2·|V |+1 ≤ ρ, otherwise decoder DECODER-HCB will be used to reconstruct the
-
TITLE WILL BE SET BY THE PUBLISHER 9
solution. We note that, if ρ = 1 the exact search algorithm will always be exe-cuted. If ρ = 0, it will never be executed and DECODER-LSQClique will behaveexactly as DECODER-IG∗. We observe that by appropriately tuning and settingthe best value for ρ one can optimize the performance of this decoder.
DECODER-LSQClique starts by creating an initial solution S′ in line 1, usingDECODER-HCB and the random keys Rj , j = 1, . . . , |V |. The loop in lines 2 to 15repeats the partial destruction (vertex eliminations) followed by the reconstruction(vertex insertions) of the current solution, until no further improvements can beobtained. The current solution S′ is copied to S in line 3. The loop in lines4 to 8 removes one by one the δ · |S′| vertices that should be eliminated fromthe current solution. A restricted candidate RCL of size max{minsize, β · |CL|}is created in line 5, containing the vertices with the smallest degrees in G(S′).The vertex with the smallest random key R|V |+j , j ∈ RCL, is selected from therestricted candidate list in line 6 and eliminated from the current solution inline 7. The reconstruction phase starts in line 9. If the random key R+2·|V |+1 issmaller than or equal to parameter ρ and G(S′) is a γ-clique, then the partialsolution S′ is extended in line 10 by an exact local search algorithm and a newsolution S∗ is obtained. Since algorithm LSQClique might stop before an optimalsolution is obtained due to its stopping criterion, then solution S∗ is completed byDECODER-HCB in line 11 (see Section 7.1 for more details). Otherwise, solutionS′ is rebuilt by DECODER-HCB in line 13, once again using the first randomkeys R+j , j = 1, . . . , |V |. The loop is interrupted in line 14 when the new solutionS′ obtained by destruction-reconstruction does not improve the incumbent S orthe graph G(S′) is not connected; otherwise a new iteration resumes. The bestsolution is returned in S.
Algorithm 3 DECODER-LSQClique(G, γ, α, δ, β,minsize, S,R+, ρ)
1: S′ ← DECODER-HCB(G, γ, α,minsize, ∅, R+j , j = 1, . . . , |V |)2: repeat3: S ← S′4: for k = 1 to δ · |S′| do5: RCL ← {v ∈ S′ : |{v′ ∈ S′ : degG(v′, S′) ≤ degG(v, S′)}| ≤
max{minsize, β · |S ′|}}6: x← argmin{R|V |+j : j ∈ RCL}7: S′ ← S′ \ {x}8: end for9: if R+2·|V |+1 ≤ ρ and dens(G(S
′)) ≥ γ then10: LSQClique(G,S′, γ, |S|, S∗)11: S′ ← DECODER-HCB(G, γ, α,minsize, S∗, R+j : j = 1, . . . , |V |)12: else13: S′ ← DECODER-HCB(G, γ, α,minsize, S′, R+j : j = 1, . . . , |V |)14: end if15: until |S′| ≤ |S| or graph G(S′) is not connected
-
10 TITLE WILL BE SET BY THE PUBLISHER
6. Exact algorithm for new decoder
The exact algorithm QClique of Ribeiro and Riveaux [24] is mainly based on aquasi-heredity proposition that states that for any γ-clique C of cardinality k ofG, there exists a permutation Π of its vertices such that, for every i = 1, . . . , k,the last vertex v of a prefix Πi of size i of Π has minimum degree in the inducedsubgraphG(Πi). In case more than one vertex has minimum degree, and supposingthe vertices are labeled 1, . . . , |V |, then one with maximum label must be the lastin the permutation. Under these two assumptions, every solution is visited onlyonce.
However, we observe that if the vertices of S′ are assumed to be those in Πi, inthis case algorithm QClique may fail to enumerate all solutions that can be built byDECODER-HCB. Consider e.g. the graph in Figure 3(a) with the current solutionS = {1, 2, 4, 7, 8} and γ = 0.7. In Figure 3(b), S is partially destroyed and thepartial solution S′ = {1, 2} is obtained. We observe that solution {1, 2, 3, 4, 5, 6}showed in Figure 3(c), can be reconstructed by DECODER-HCB, but not byalgorithm QClique starting with Πi =< 1, 2 >, since vertex 2 of minimum degreein G(S′) would not be the last one in all possible permutations with prefix Πi.
5
6
3
4
1
2
7
8
5
6
3
4
1
2
7
8
(a) (b)
5
6
3
4
1
2
7
8
(c)
Figure 3. Maximum solution {1, 2, 3, 4, 5, 6} can not be recon-structed by QClique from the initial solution S′ = {1, 2} after thepartial destruction of S = {1, 2, 4, 7, 8}
Algorithm LSQClique is used as a variant of algorithm QClique in line 10 ofAlgorithm 3 to perform the exact local search. It does not make use of the quasi-heredity proposition [24]. It may investigate multiple permutations of the samesolution along the search tree, contrarily to QClique that visits only one.
Instead of working with permutations like in the original algorithm QClique,algorithm LSQclique works directly with the initial solution represented by theset of vertices S′. The partial solution S′ and the initial lower bound LB = |S|
-
TITLE WILL BE SET BY THE PUBLISHER 11
Algorithm 4 LSQClique(G,S′, γ,LB , S∗)
1: S∗ ← S′2: i← |S′|3: if a solution of size k containing all vertices of S′ does not exist for somek = i+ 1, . . . ,LB + 1 then return
4: CL← {v ∈ V \ S′ : dens(G(S′ ∪ {v})) ≥ γ}5: if CL = ∅ then return6: for all j ∈ CL do7: LSQClique(G,S′ ∪ {j}, γ,max{LB, |S′|+ 1}, S′′)8: if |S′′| > |S∗| then9: S∗ ← S′′
10: LB ← max{LB, |S∗|}11: end if12: end for
are passed to Algorithm 4 as parameters. Line 1 of the pseudo-code copies to S∗the current partial solution S′. Line 2 sets i with the size of the current solutionS′. If a solution of size k containing all vertices in S′ does not exist for some valueof k = i+ 1, . . . , LB + 1, then the algorithm returns in line 3 (pruning).
The candidate list CL created in line 4 is formed by every vertex v ∈ V \ S′ :dens(G(S′ ∪ {v})) ≥ γ, as for DECODER-HCB. If the candidate list is empty,then the algorithm returns in line 5.
The loop in lines 6 to 12 attempts to extend the current solution S′ by eachvertex j in the candidate list CL. The vertices in the candidate list are takenin non increasing order of their degrees degG(j, S′). For each of them, line 7recursively invokes the algorithm for a new, extended partial solution S′∪{j} anda new, updated lower bound max{LB, |S′|+1}, returning a new solution S′′. Line8 checks if the new solution S′′ improves then current best S∗. If this is the case,the current best and the lower bound are updated in lines 9 and 10, respectively.When the recursion terminates, S∗ contains the best solution obtained by thealgorithm.
The pruning strategy discards the current node of the enumeration tree in line3 if a solution S of size k containing all vertices of S′ can not exist for somek = i+ 1, . . . , LB + 1. This is done by calculating an upper bound to the numberof edges in G(S) and checking if the number of edges in a γ-clique is greater thanthis upper bound. We follow a similar idea to that in [24].
Let E(S′, S \ S′) denote the set of edges in G between the vertices of S′ andS \ S′. If S is a γ-clique, then the following condition holds and gives a lowerbound to |E(S)|:
|E(S)| = |E(S′) ∪ E(S′, S \ S′) ∪ E(S \ S′)| ≥ γ ·(k
2
). (8)
-
12 TITLE WILL BE SET BY THE PUBLISHER
Since the sets in Equation (8) are disjoint,
|E(S)| = |E(S′)|+ |E(S′, S \ S′)|+ |E(S \ S′)|. (9)
Since |E(S′)| is already known, in order to calculate an upper bound to |E(S)|we need to calculate an upper bound to |E(S′, S \ S′)|+ |E(S \ S′)|.
Let V ′ be formed by the k − i vertices of V \ S′ with the maximum numberof adjacent vertices in G that belong to S′. V ′ can be determined in polynomialtime O(|V |): first sort the vertices in the non increasing order of their degreesdegG(v, S), then select the k− i first vertices. Then, an upper bound to |E(S′, S \S′)| is given by |E(S′, V ′)|.
Now, let us define V ′′ as the set formed by the k − i vertices of V \ S′ withmaximum degree in G. Then, the following condition holds:∑
v∈S\S′degG(v) ≤
∑v∈V ′′
degG(v). (10)
Following a reasoning similar to that in [24], we obtain:
|E(S′, S \ S′)|+ |E(S \ S′)| ≤∑v∈V ′
degG(v, S′) + min
{∑v∈V ′′ degG(v)−
∑v∈V ′ degG(v, S
′)
2,
(k − i
2
)}.
(11)
Therefore, if the inequality below is not satisfied for at least one value of k =i+ 1, . . . , LB + 1, then a solution of size LB + 1 does not exist and we can prunethe current node of the search tree:
|E(S′)|+∑v∈V ′
degG(v, S′) + min
{∑v∈V ′′ degG(v)−
∑v∈V ′ degG(v, S
′)
2,
(k − i
2
)}≥ γ ·
(k
2
).
(12)
7. Computational results
7.1. Experimental setting
All algorithms were implemented using version 19.14.26429.4 of the Microsoft(R) C/C++ Optimizing Compiler for x64. The computational experiments havebeen performed on an Intel Core i5-5200 processor with 2.20 GHz and 8 GB ofRAM running under Windows 10.
-
TITLE WILL BE SET BY THE PUBLISHER 13
Table 1. Parameters settings.
parameter BRKGA-IG∗ BRKGA-LSQCliquep 91 91pe 0.13 0.13pm 0.22 0.22rhoe 0.78 0.78α 0.01 0.01
δ 0.40 1− dens(G); if dens(G) < 0.82 · (1− dens(G)); if dens(G) ≥ 0.8
β 0.02 0.02maxnodes - 220ρ - max{1− dens(G), 0.80}
The newly proposed algorithm BRKGA-LSQClique was compared with theoriginal BRKGA-IG∗ heuristic of [22], which was the best heuristic for the maxi-mum quasi-clique problem at the time of writing. The best parameters for algo-rithm BRKGA-IG∗ were determined using the automatic tuning tool IRACE [16,21]. The same parameter settings were used for algorithm BRKGA-LSQClique,except for the parameter δ that controls the number of vertices to be destroyed. Inorder to enforce that the fraction of nodes eliminated in the destruction phase besmaller for dense graphs, DECODER-LSQClique empirically sets δ = 1−dens(G),if dens(G) < 0.8; δ = 2 · (1− dens(G)), otherwise.
DECODER-LSQClique makes use of an additional parameter ρ to establishwhether the exact local search should be applied or not to some solution. Again,we empirically set the probability that exact local search is applied at ρ = max{1−dens(G), 0.80}.
An additional parameter maxnodes is used to enforce a maximum limit to thenumber of nodes of the search tree generated by the exact algorithm LSQClique.It was set at 220 using the IRACE tool. In case the search tree generated byLSQClique reaches this maximum limit of nodes, then the algorithm returns theincumbent solution. For this reason, an additional application of the reconstruc-tion procedure DECODER-HCB was enforced in line 11 of Algorithm 3 to furtherimprove the current solution.
All parameter settings are shown in Table 1.We considered 27 DIMACS instances [14], 33 maximum clique instances of the
Benchmarks with Hidden Optimum Solutions for Graph Problems (BHOSLIB) [5,25, 26], and ten sparse instances from the University of Florida Sparse MatrixCollection [9] in the computational experiments.
7.2. Numerical results
In the first experiment, for each instance and each value of the threshold γ, weperformed 30 runs of each algorithm BRKGA-IG∗ and BRKGA-LSQClique until atarget solution value was found. Each run was limited to ten minutes of execution,
-
14 TITLE WILL BE SET BY THE PUBLISHER
except for the instances on Table 7, whose runs were limited to 30 minutes ofexecution. We discarded the instances for which both algorithm obtained thetarget solution value for every value of γ in less than three generations on averageover the 30 runs.
Tables 2 to 7 show the results for 13 of the 27 DIMACS instances, all 33BHOSLIB instances, and three of the ten sparse instances that have not beendiscarded. For each instance, the tables shows its density, the threshold γ, the tar-get look4 considered in the experiment, and the results obtained by each algorithmBRKGA-IG∗ and BRKGA-LSQClique. For each algorithm, the table presents theaverage and best sizes of the solution found over the 30 runs, the average runningtime in seconds and the average generation in which the target was found over theruns that reached the target value (#look4). Whenever any of the two methodsfailed to find a solution as good as the target value look4 within the time limitfor all instances in the group of same density, we indicate it by “> 600 (0)” or“> 1800 (0)”.
The rows highlighted in dark gray ( ) correspond to the cases where BRKGA-IG∗ reached the target look4 more often or, when both of them reached the targetthe same number of times, it was faster than BRKGA-LSQClique in terms of theaverage time to target. Rows in light gray ( ) are those for which BRKGA-LSQClique performed better in terms of the same criteria. For all other cases,both algorithms found the target in the first generation in all runs.
These results show that BRKGA-IG∗ performed better for 43 out of the 46instances with density greater than 0.5 and tested with γ = 0.999, while BRKGAperfomed better for the remaining three (DSJC1000.5, frb30-15-1, and C4000.5).However, for smaller values of the threhold γ, algorithm BRKGA-LSQClique con-siderably outperformed BRKGA-IG∗ for 47 instances (i.e., for all but instancesbrock400_2 and PGPgiantcompo), showing that the exact local search is capa-ble to considerably improve the performance of the biased random key geneticalgorithm when the target is smaller than 0.999.
In the next experiment, we evaluate and compare the run time distributions(or time-to-target plots) of algorithms BRKGA-IG∗ and BRKGA-LSQClique forsome instances. Time-to-target plots display on the ordinate axis the probabilitythat an algorithm will find a solution at least as good as a given target valuewithin a given running time, shown on the abscissa axis. Run time distributionshave also been advocated by Hoos and Stützle [13] as a way to characterize therunning times of stochastic local search algorithms for combinatorial optimizationproblems. In this experiment, the two algorithms were made to stop whenever asolution with cost greater than or equal to a given target value was found. Thetargets are the same used in the first experiment. Each heuristic was run 200 timesfor each value of the threshold γ. Next, the empirical probability distributions ofthe time taken by each heuristic to find the target solution value are plotted. Toplot the empirical distribution for each heuristic, we followed the methodologyproposed by Aiex et al. [2, 3]. We associate a probability pi = (i − 12 )/200 withthe i-th smallest running time ti and plot the points (ti, pi), for i = 1, . . . , 200.The more to the left is a plot, the better is the algorithm corresponding to it.
-
TITLE WILL BE SET BY THE PUBLISHER 15
BRKGA-IG*
BRKGA-LSQ
Cliq
ueInstan
ceDens.
γlo
ok4
Avg
.Best
Tim
e(#
look
4)Gen.
Avg
.Best
Tim
e(#
look
4)Gen.
C250.9
0.899
0.999
4444.00
440.18
(30)
2.77
44.00
440.40
(30)
4.37
0.95
105
105.00
105
0.52
(30)
2.40
105.03
106
0.43
(30)
1.07
C500.9
0.901
0.999
5757.00
5710.79(30)
49.40
57.00
5712.15(30)
55.10
0.95
157
157.00
157
8.72
(30)
10.17
157.00
157
4.19
(30)
4.83
C1000.9
0.901
0.999
6767.10
68126.51
(30)
60.97
67.00
68170.53
(27)
103.26
0.95
217
217.30
219
128.51
(30)
27.30
217.43
219
36.15(27)
6.37
DSJ
C500.5
0.500
0.999
1313.00
130.38
(30)
1.67
13.00
130.61
(30)
1.53
0.9
2121.00
213.04
(30)
4.87
21.00
212.89
(30)
3.10
0.8
3434.00
3420.66(30)
21.70
34.00
3410.85(30)
10.13
DSJ
C1000.5
0.502
0.999
1514.67
15265.43
(20)
349.20
14.73
15180.49
(22)
180.36
0.9
2423.30
24254.91
(9)
82.22
23.63
24203.97
(19)
71.05
0.8
3938.87
40242.46
(27)
31.59
39.03
4041.22(30)
7.03
brock4
00_1
0.748
0.999
2525.00
25101.51
(30)
852.43
24.83
25294.50
(25)
1504.56
0.9
6060.00
6023.60(30)
47.30
60.00
6010.78(30)
14.37
0.8
189
189.00
189
2.62
(30)
3.10
189.00
189
1.84
(30)
1.27
brock4
00_2
0.749
0.999
2525.00
251.57
(30)
13.10
25.00
254.81
(30)
22.43
0.9
6161.00
611.88
(30)
3.67
61.00
612.30
(30)
3.23
0.8
186
186.00
186
4.50
(30)
6.00
186.00
186
1.74
(30)
1.63
brock4
00_3
0.748
0.999
2525.00
251.25
(30)
10.47
25.00
251.57
(30)
9.10
0.9
6161.00
6129.52(30)
56.03
61.00
6114.58(30)
20.93
0.8
187
187.00
187
29.114
(30)
39.60
187.00
187
6.09
(30)
2.47
brock8
00_1
0.649
0.999
2121.00
2150.07(30)
108.77
21.00
2191.67(30)
135.43
0.9
4241.87
42143.39
(26)
54.88
42.00
4268.18(30)
26.67
0.8
9695.77
96254.22
(23)
80.22
96.00
9658.27(30)
19.03
brock8
00_2
0.651
0.999
2121.00
2154.91(30)
106.90
21.00
21131.81
(30)
189.13
0.9
4241.93
42182.78
(28)
70.07
42.00
4398.38(29)
39.59
0.8
9494.03
9594.59(30)
30.87
94.00
9433.99(30)
11.27
Tabl
e2.
First
expe
riment:
30runs
ofeach
instan
cean
dalgorithm
limited
to600second
sof
runn
ingtimeeach.
-
16 TITLE WILL BE SET BY THE PUBLISHER
BRKGA-IG*
BRKGA-LSQ
Cliq
ueInstan
ceDens.
γlo
ok4
Avg
.Best
Tim
e(#
look
4)Gen.
Avg
.Best
Tim
e(#
look
4)Gen.
brock8
00_3
0.649
0.999
2121.00
2111.83(30)
18.67
21.00
2115.97(30)
24.17
0.9
4241.97
42190.77
(29)
79.76
42.00
4228.47(30)
12.30
0.8
9392.73
93237.31
(22)
78.91
93.00
9359.67(30)
20.27
frb3
0-15-1
0.824
0.999
2828.13
291.95
(30)
13.26
28.10
291.88
(30)
9.43
0.95
5959.00
597.31
(30)
13.07
59.00
597.21
(30)
9.80
0.90
139
139.00
139
25.30(30)
31.47
139.00
138
12.02(30)
11.36
frb3
0-15-2
0.823
0.999
2828.03
290.82
(30)
4.96
28.03
291.20
(30)
5.06
0.95
5858.00
5811.38(30)
19.40
58.00
588.03
(30)
10.60
0.90
137
137.40
138
281.27
(30)
24.60
312.58
138
220.50
(30)
195.72
frb3
0-15-4
0.824
0.999
2929.00
2958.10(30)
394.97
29.00
2972.97(30)
332.63
0.95
6160.80
61243.11
(24)
364.58
60.80
61169.83
(24)
183.75
0.90
138
137.70
138
255.28
(21)
303.10
137.97
138
163.85
(29)
150.90
frb3
0-15-5
0.824
0.999
2929.00
2955.16(30)
375.47
29.00
29137.72
(30)
694.60
0.95
6060.00
6099.01(30)
157.43
60.00
6042.16(30)
46.63
0.90
134
134.00
134
19.61(30)
24.60
134.00
134
11.10(30)
11.57
frb3
5-17-1
0.842
0.999
3333.00
3341.66(30)
168.80
33.00
33108.88
(30)
372.70
0.95
7877.97
78112.96
(29)
92.83
78.00
7867.17(30)
47.77
0.90
207
215.20
216
1.81
(30)
1.00
215.47
216
1.89
(30)
1.00
frb3
5-17-2
0.842
0.999
3333.00
3321.08(30)
85.87
33.00
3332.79(30)
115.53
0.95
7575.10
7681.34(30)
63.47
75.10
7633.02(30)
23.03
0.90
207
207.00
207
100.84
(30)
60.60
207.00
207
43.63(30)
23.40
frb3
5-17-4
0.843
0.999
3333.00
3347.44(30)
187.57
33.00
3377.14(30)
247.93
0.95
7979.07
8023.49(30)
18.93
79.17
8113.27(30)
9.27
0.90
238
237.77
238
258.20
(23)
170.00
238.00
238
155.42
(30)
98.00
frb3
5-17-5
0.841
0.999
3333.03
3415.21(30)
60.90
33.00
3324.78(30)
87.80
0.95
7979.00
7986.15(30)
69.50
79.03
8033.77(30)
23.07
0.90
225
224.53
225
389.49
(16)
252.63
225.00
225
109.48
(30)
71.33
Tabl
e3.
First
expe
riment:
30runs
ofeach
instan
cean
dalgorithm
limited
to600second
sof
runn
ingtimeeach.
-
TITLE WILL BE SET BY THE PUBLISHER 17
BRKGA-IG*
BRKGA-LSQ
Cliq
ueInstan
ceDens.
γlo
ok4
Avg
.Best
Tim
e(#
look
4)Gen.
Avg
.Best
Tim
e(#
look
4)Gen.
frb4
0-19-1
0.857
0.999
3737.00
3727.10(30)
68.90
37.03
3832.11(30)
73.73
0.95
111
111.03
112
53.17(30)
22.30
111.20
113
27.26(30)
9.77
0.90
435
435.00
435
2.23
(30)
1.03
435.00
435
2.03
(30)
1.00
frb4
0-19-2
0.857
0.999
3737.00
3715.78(30)
40.13
37.00
3723.98(30)
55.23
0.95
103
102.90
104
178.76
(26)
82.04
103.07
104
71.23(30)
30.57
0.90
363
363.00
363
16.89(30)
6.80
363.00
363
6.77
(30)
2.83
frb4
0-19-4
0.856
0.999
3837.83
39145.57
(24)
370.67
37.60
38252.25
(18)
573.72
0.95
9595.00
9553.74(30)
24.57
95.20
9622.56(30)
9.33
0.90
292
291.10
292
395.01
(4)
160.50
291.73
292
229.11
(22)
90.50
frb4
0-19-5
0.856
0.999
3838.00
3899.55(30)
254.43
37.97
38120.16
(29)
282.14
0.95
100
99.50
100
264.97
(15)
115.07
99.97
100
123.23
(29)
51.97
0.90
333
332.80
333
219.13
(24)
87.38
333.00
333
32.61(30)
13.17
frb4
5-21-1
0.857
0.999
4141.07
4212.20(30)
18.97
41.03
4226.03(30)
37.90
0.95
121
121.33
123
43.20(30)
10.37
121.57
124
15.65(30)
3.63
0.90
508
508.00
508
35.54(30)
10.27
508.00
508
7.10
(30)
2.20
frb4
5-21-2
0.869
0.999
4141.07
4215.26(30)
23.63
41.03
4241.90(30)
62.47
0.95
120
120.23
122
71.53(30)
16.93
120.33
122
39.44(30)
9.83
0.90
491
490.17
491
291.03
(5)
73.80
491.00
491
151.37
(30)
42.80
frb4
5-21-4
0.869
0.999
4242.00
42125.88
(30)
189.07
41.93
42183.97
(28)
260.39
0.95
128
128.00
128
126.04
(30)
30.83
128.33
130
43.94(30)
9.67
0.90
554
554.00
554
17.76(30)
5.00
554.00
554
3.44
(30)
1.00
frb4
5-21-5
0.869
0.999
4241.90
43195.14
(26)
302.19
41.57
42186.09
(17)
274.47
0.95
121
120.93
123
177.89
(27)
44.82
121.20
122
96.31(30)
23.33
0.90
508
507.13
508
176.41
(4)
43.75
508.00
508
61.36(30)
16.97
frb5
0-23-1
0.879
0.999
4646.03
4729.85(30)
15.97
46.03
4770.75(30)
38.87
0.95
157
157.30
159
295.15
(28)
28.46
157.53
160
54.86(30)
6.63
0.90
797
796.03
797
87.44(1)
11.00
797.00
797
52.13(30)
8.13
Tabl
e4.
First
expe
riment:
30runs
ofeach
instan
cean
dalgorithm
limited
to600second
sof
runn
ingtimeeach.
-
18 TITLE WILL BE SET BY THE PUBLISHER
BRKGA-IG*
BRKGA-LSQ
Cliq
ueInstan
ceDens.
γlo
ok4
Avg
.Best
Tim
e(#
look
4)Gen.
Avg
.Best
Tim
e(#
look
4)Gen.
frb5
0-23-2
0.878
0.999
4646.03
4727.28(30)
14.53
46.00
4661.16(30)
33.60
0.95
158
156.13
158
397.67
(6)
40.17
158.13
159
193.64
(30)
24.40
0.90
780
779.93
780
154.18
(28)
20.07
780.00
780
18.71(30)
2.80
frb5
0-23-4
0.879
0.999
4747.00
47159.94
(30)
85.43
46.77
47185.04
(23)
99.87
0.95
157
156.43
158
350.15
(19)
36.11
157.40
160
154.65
(29)
20.97
0.90
772
771.00
771
>600(0)
-771.60
772
217.30
(18)
33.61
frb5
0-23-5
0.879
0.999
4646.10
4726.44(30)
13.60
46.13
4754.29(30)
29.43
0.95
160
158.53
162
356.84
(6)
33.50
160.43
162
201.76
(30)
26.43
0.90
788
788.00
788
77.23(30)
10.00
788.00
788
8.17
(30)
1.17
frb5
3-24-1
0.883
0.999
5049.10
50366.36
(6)
123.83
48.77
49>600(0)
-0.95
199
198.57
200
336.91
(17)
20.88
199.50
201
142.60
(30)
10.93
0.90
978
977.43
978
272.78
(13)
27.46
978.00
978
23.79(30)
2.63
frb5
3-24-2
0.883
0.999
5049.43
51189.84
(13)
64.38
49.30
50216.89
(12)
77.08
0.95
172
169.80
172
382.90
(5)
26.20
172.07
174
202.27
(29)
18.00
0.90
926
925.07
926
285.68
(2)
25.00
925.90
926
123.69
(27)
13.00
frb5
3-24-4
0.883
0.999
5049.00
50283.98
(5)
91.40
48.70
50317.79
(4)
107.75
0.95
178
177.83
179
298.29
(22)
19.18
178.40
180
78.62(30)
5.80
0.90
932
932.00
932
13.05(30)
1.00
932.00
932
10.51(30)
1.00
frb5
3-24-5
0.883
0.999
4948.97
49165.88
(29)
54.83
48.80
50194.90
(23)
65.26
0.95
166
166.50
168
302.86
(26)
20.58
167.13
171
69.86(30)
5.50
0.90
894
893.00
893
>600(0)
-894.00
894
125.81
(30)
11.70
frb5
6-25-1
0.888
0.999
5351.63
53236.63
(4)
66.00
51.20
53188.83
(1)
56.00
0.95
225
224.30
225
242.23
(15)
10.80
225.33
227
69.21(30)
3.27
0.90
1161
1161.00
1161
52.68(30)
4.27
1161.00
1161
11.05(30)
1.00
frb5
6-25-2
0.888
0.999
5251.43
53280.08
(12)
74.17
51.03
52152.36
(4)
42.75
0.95
213
210.63
212
>600(0)
-213.80
217
198.41
(29)
12.00
0.90
1175
1175.00
1175
12.94(30)
1.00
1175.00
1175
10.68(30)
1.00
Tabl
e5.
First
expe
riment:
30runs
ofeach
instan
cean
dalgorithm
limited
to600second
sof
runn
ingtimeeach.
-
TITLE WILL BE SET BY THE PUBLISHER 19
BRKGA-IG*
BRKGA-LSQ
Cliq
ueInstan
ceDens.
γlo
ok4
Avg
.Best
Tim
e(#
look
4)Gen.
Avg
.Best
Tim
e(#
look
4)Gen.
frb5
6-25-4
0.888
0.999
5151.07
5292.58(30)
25.20
51.00
52131.59
(29)
37.00
0.95
195
193.33
196
221.82
(1)
10.00
195.33
197
128.24
(30)
7.47
0.90
1125
1125.00
1125
16.50(30)
1.10
1125.00
1125
12.32(30)
1.00
frb5
6-25-5
0.888
0.999
5251.53
53240.08
(15)
66.40
51.23
52165.91
(8)
45.50
0.95
201
199.10
202
515.94
(6)
23.33
201.50
203
188.53
(30)
11.41
0.90
1138
1138.00
1138
15.19(30)
1.03
1138.00
1138
12.30(30)
1.00
frb5
9-26-1
0.892
0.999
5554.17
56346.35
(7)
81.00
53.57
55242.13
(2)
55.50
0.95
240
239.03
241
318.16
(10)
11.30
240.90
244
156.25
(30)
6.83
0.90
1388
1388.00
1388
15.50(30)
1.00
1388.00
1388
11.28(30)
1.00
frb5
9-26-2
0.893
0.999
5453.83
54199.91
(25)
46.44
53.60
54201.10
(18)
49.06
0.95
235
234.57
237
351.97
(16)
13.25
235.80
238
102.22
(30)
4.367
0.90
1388
1388.00
1388
13.60(30)
1.00
1388.00
1388
11.10(30)
1.00
frb5
9-26-4
0.892
0.999
5554.33
55228.03
(11)
52.45
53.73
55322.46
(4)
78.50
0.95
230
228.27
232
340.83
(6)
13.50
230.70
233
203.91
(30)
10.87
0.90
1373
1373.00
1373
14.22(30)
1.00
1373.00
1373
11.42(30)
1.00
frb5
9-26-5
0.893
0.999
5453.80
55198.84
(22)
46.32
53.50
54282.46
(15)
68.40
0.95
221
220.83
223
297.89
(21)
13.91
221.90
224
89.80(30)
3.73
0.90
1360
1360.00
1360
15.62(30)
1.00
1360.00
1360
12.46(30)
1.00
Tabl
e6.
First
expe
riment:
30runs
ofeach
instan
cean
dalgorithm
limited
to600second
sof
runn
ingtimeeach.
-
20 TITLE WILL BE SET BY THE PUBLISHER
BRKGA-IG*
BRKGA-LSQ
Cliq
ueInstan
ceDens.
γlo
ok4
Avg
.Best
Tim
e(#
look
4)Gen.
Avg
.Best
Tim
e(#
look
4)Gen.
C2000.9
0.90
0.999
7473.70
7525.85(20)
247.92
73.47
7425.70(15)
305.73
0.95
276
270.50
275
>1800
(0)
-276.30
278
349.44
(25)
10.56
C4000.5
0.50
0.999
1716.93
17461.87
(28)
19.00
16.97
18353.46
(28)
14.79
0.8
4843.60
47>1800
(0)
-49.76
49449.38
(30)
5.35
0.90
2726.83
27761.12
(25)
6.76
27.07
28115.73
(30)
1.20
frb1
00-40
0.93
0.999
8685.90
87779.23
(23)
13.04
85.03
86771.79
(8)
16.00
0.95
1835
1834.97
1835
885.90
(28)
5.286
1836.00
1837
322.2(30)
1.53
EVA
0.00019
0.90
44.00
41.67
(30)
6.30
4.00
41.82
(30)
2.73
0.80
44.57
50.32
(30)
1.00
4.97
50.61
(30)
1.00
0.70
55.00
50.34
(30)
1.00
5.00
50.65
(30)
1.00
0.60
77.00
73.52
(30)
3.17
7.00
72.21
(30)
1.10
0.50
99.00
92.31
(30)
1.57
9.00
92.28
(30)
1.00
Geom
0.00044
0.90
2323.00
233.99
(30)
8.17
23.00
231.65
(30)
1.00
0.80
2325.90
262.32
(30)
1.00
26.00
263.38
(30)
1.00
0.70
3030.00
3038.06(30)
3.53
30.00
307.25
(30)
1.13
0.60
3635.53
36607.54
(16)
32.00
36.00
36420.33
(30)
28.70
0.50
4443.80
44486.85
(24)
17.79
44.00
4418.50(30)
1.00
PGPgian
t-0.00043
0.90
4242.00
423.77
(30)
1.10
42.00
425.07
(30)
1.10
compo
0.80
4848.00
4825.23(30)
2.40
48.00
489.31
(30)
1.03
0.70
5252.00
5212.63(30)
1.10
52.00
5214.14(30)
1.03
0.60
6060.00
6050.92(30)
2.23
60.00
6021.59(30)
1.07
0.50
7373.00
73173.76
(30)
4.33
73.00
7326.58(30)
1.00
Tabl
e7.
First
expe
riment:
30runs
ofeach
instan
cean
dalgorithm
limited
to1800
second
sof
runn
ingtimeeach.
-
TITLE WILL BE SET BY THE PUBLISHER 21
Figures 4 and 5 illustrate the time-to-target plots for instance brock400_2 withγ = 0.999, 0.90, and 0.80 and for instance PGPgiantcompo with γ = 0.90, 0.80,and 0.50, respectively. We recall from Tables 2 and 7 that BRKGA-IG∗ obtainedbetter results for both instances with γ = 0.90 and for PGPgiantcompo withγ = 0.70. The plots in these figures show that BRKGA-IG∗ performs better forthe highest values of γ, while BRKGA-LSQClique becomes progressively better asγ decreases.
8. Concluding remarks
In this work, we showed that the exact enumeration algorithmQClique proposedby Ribeiro and Riveaux [24] can be hybridized with the biased random-key geneticalgorithm BRKGA-IG∗ developed by Pinto et al. [22] as a local search strategyto improve the quality of the solutions created by the decoder. The new decoderusing an exact local search strategy gives rise to a matheuristic for solving themaximum cardinality quasi-clique problem based on a biased random-key geneticalgorithm.
Computational experiments were performed on a set of 70 test instances con-sidering different thresholds.
The results showed that although BRKGA-IG∗ performed better for 43 out ofthe 46 instances with density greater than 0.5 and tested with γ = 0.999, algorithmBRKGA-LSQClique becomes progressively better as the threshold γ decreases.For smaller values of γ, algorithm BRKGA-LSQClique considerably outperformedBRKGA-IG∗ for 47 instances, showing that the exact local search is capable toconsiderably improve the performance of the biased random key genetic algorithmwhen the target is smaller than 0.999.
References
[1] J. Abello, M. Resende, and S. Sudarsky. Massive quasi-clique detection. InJ. Abello and J. Vitter, editors, Proceedings of the 5th Latin American Sym-posium on the Theory of Informatics, pages 598–612. Springer, 2002.
[2] R. Aiex, M. Resende, and C. C. Ribeiro. Probability distribution of solutiontime in GRASP: An experimental investigation. Journal of Heuristics, 8:343–373, 2002.
[3] R. Aiex, M. Resende, and C. C. Ribeiro. TTTPLOTS: A Perl program tocreate time-to-target plots. Optimization Letters, 1:355–366, 2007.
[4] J. C. Bean. Genetic algorithms and random keys for sequencing and opti-mization. ORSA Journal on Computing, 2:154–160, 1994.
[5] BHOSLIB. Benchmarks with hidden optimum solutions for graph problems.http://networkrepository.com/, 2004. Online reference, last visited on June7, 2018.
-
22 TITLE WILL BE SET BY THE PUBLISHER
[6] J. S. Brandão, T. F. Noronha, M. G. C. Resende, and C. C. Ribeiro. A bi-ased random-key genetic algorithm for single-round divisible load scheduling.International Transactions Operational Research, 22:823–839, 2015.
[7] J. S. Brandão, T. F. Noronha, M. G. C. Resende, and C. C. Ribeiro. A biasedrandom-key genetic algorithm for scheduling heterogeneous multi-round sys-tems. International Transactions Operational Research, 27:1061–1077, 2017.
[8] M. Brunato, H. Hoos, and R. Battiti. On effectively finding maximal quasi-cliques in graphs. In V. Maniezzo, R. Battiti, and J.-P. Watson, editors,Learning and Intelligent Optimization, volume 5313 of Lecture Notes in Com-puter Science, pages 41–55. Springer, Berlin, 2008.
[9] T. A. Davis and Y. Hu. The University of Florida sparse matrix collection.ACM Transactions on Mathematical Software, 38(1), 2011.
[10] FICO. FICO Xpress Optimization Suite 7.6.http://www.fico.com/en/products/fico-xpress-optimization-suite, 2017.Online reference, last visited on May 25, 2017.
[11] J. F. Goncalves and M. G. C. Resende. Biased random-key genetic algorithmsfor combinatorial optimization. Journal of Heuristics, 17:487–525, 2011.
[12] J. F. Goncalves, M. G. C. Resende, and R. F. Toso. Biased and unbiasedrandom key genetic algorithms: An experimental analysis. In Abstracts of the10th Metaheuristics International Conference, Singapore, 2013.
[13] H. Hoos and T. Stützle. Evaluation of Las Vegas algorithms - Pitfalls andremedies. In G. Cooper and S. Moral, editors, Proceedings of the 14th Confer-ence on Uncertainty in Artificial Intelligence, pages 238–245, Madison, 1998.
[14] D. S. Johnson, editor. Cliques, Coloring, and Satisfiability: Second DIMACSImplementation Challenge, volume 26 of Dimacs Series in Discrete Mathe-matics and Theoretical Computer Science. American Mathematical Society,Providence, 1996.
[15] R. M. Karp. Reducibility among combinatorial problems. In R. E. Millerand J. W. Thatcher, editors, Complexity of Computer Computations, pages85–103, New York, 1972. Plenum.
[16] M. López-Ibánez, J. Dubois-Lacoste, T. Stützle, and M. Birattari. TheIRACE package: Iterated race for automatic algorithm configuration, 2011.Technical Report TR/IRIDIA/2011-004, IRIDIA, Université Libre de Brux-elles, Belgium.
[17] T. F. Noronha, M. G. C. Resende, and C. C. Ribeiro. A biased random-keygenetic algorithm for routing and wavelength assignment. Journal of GlobalOptimization, 50:503–518, 2011.
[18] A. B. Oliveira, A. Plastino, and C. C. Ribeiro. Construction heuristics forthe maximum cardinality quasi-clique problem. In Abstracts of the 10th Meta-heuristics International Conference, page 84, Singapore, 2013.
[19] F. M. Pajouh, Z. Miao, and B. Balasundaram. A branch-and-bound ap-proach for maximum quasi-cliques. Annals of Operations Research, 216:145–161, 2014.
[20] J. Pattillo, A. Veremyev, S. Butenko, and V. Boginski. On the maximumquasi-clique problem. Discrete Applied Mathematics, 161:244–257, 2013.
-
TITLE WILL BE SET BY THE PUBLISHER 23
[21] L. Pérez Cáceres, M. López-Ibáñez, and T. Stützle. An analysis of parametersof IRACE. In Proceedings of the 14th European Conference on EvolutionaryComputation in Combinatorial Optimization, volume 8600 of Lecture Notesin Computer Science, pages 37–48. Springer, Berlin, 2014.
[22] B. Q. Pinto, C. C. Ribeiro, I. Rosseti, and A. Plastino. A biased random-keygenetic algorithm for solving the maximum quasi-clique problem. EuropeanJournal of Operational Research, 271:849–865, 2018.
[23] M. G. C. Resende and C. C. Ribeiro. Biased-random key genetic algorithms:An advanced tutorial. In Proceedings of the 2016 Genetic and EvolutionaryComputation Conference - GECCO’16 Companion Volume, pages 483–514,Denver, 2016. Association for Computing Machinery.
[24] C. C. Ribeiro and J. A. Riveaux. An exact algorithm for the maximum quasi-clique problem. International Transactions in Operational Research, 26, 2019.
[25] R. A. Rossi and N. K. Ahmed. Coloring large complex networks. SocialNetwork Analysis and Mining, 4(1), 2014.
[26] R. A. Rossi and N. K. Ahmed. The network data repository with interactivegraph analytics and visualization. In Proceedings of the Twenty-Ninth AAAIConference on Artificial Intelligence, pages 4292–4293, 2015.
[27] W. Spears and K. de Jong. On the virtues of parameterized uniform crossover.In R. Belew and L. Booker, editors, Proceedings of the Fourth InternationalConference on Genetic Algorithms, pages 230–236, San Mateo, 1991. MorganKaufman.
[28] R. F. Toso and M. G. C. Resende. A C++ application programming inter-face for biased random-key genetic algorithms. Optimization Methods andSoftware, 30:81–93, 2015.
[29] A. Veremyev, O. A. Prokopyev, S. Butenko, and E. L. Pasiliao. Exact MIP-based approaches for finding maximum quasi-clique and dense subgraph.Computational Optimization and Applications, 64:177–214, 2016.
-
24 TITLE WILL BE SET BY THE PUBLISHER
Figure 4. Time-to-target plot for instance brock400_2.
0
0.2
0.4
0.6
0.8
1
0 5 10 15 20 25 30
cu
mu
lative
pro
ba
bili
ty
time to target solution
BRKGA-LSQCliqueBRKGA-IG*
(a) γ = 0.999
0
0.2
0.4
0.6
0.8
1
0 2 4 6 8 10 12 14 16
cu
mu
lative
pro
ba
bili
ty
time to target solution
BRKGA-LSQCliqueBRKGA-IG*
(b) γ = 0.90
0
0.2
0.4
0.6
0.8
1
0 5 10 15 20 25 30
cu
mu
lative
pro
ba
bili
ty
time to target solution
BRKGA-LSQCliqueBRKGA-IG*
(c) γ = 0.80
-
TITLE WILL BE SET BY THE PUBLISHER 25
Figure 5. Time-to-target plot for instance PGPgiantcompo.
0
0.2
0.4
0.6
0.8
1
0 5 10 15 20 25 30 35
cu
mu
lative
pro
ba
bili
ty
time to target solution
BRKGA-LSQCliqueBRKGA-IG*
(a) γ = 0.90
0
0.2
0.4
0.6
0.8
1
0 50 100 150 200 250 300
cu
mu
lative
pro
ba
bili
ty
time to target solution
BRKGA-LSQCliqueBRKGA-IG*
(b) γ = 0.80
0
0.2
0.4
0.6
0.8
1
0 100 200 300 400 500 600
cu
mu
lative
pro
ba
bili
ty
time to target solution
BRKGA-LSQCliqueBRKGA-IG*
(c) γ = 0.50