chapter 3: evolutionary fuzzy...
TRANSCRIPT
Chapter 3: Evolutionary Fuzzy Modeling
73
A Genetic-Fuzzy Approach to Measure Multiple Intelligence
Chapter 3: Evolutionary Fuzzy Modeling
Genetic- Fuzzy Systems(GFS)
Machine Learning Models
Advantages
Limitations
Applications
Need of Research Framework
Chapter 3: Evolutionary Fuzzy Modeling
74
A Genetic-Fuzzy Approach to Measure Multiple Intelligence
3.1 Chapter Overview
This chapter presents significant design views of genetic learning. In order to achieve
hybridization of Fuzzy Logic (FL) with Genetic Algorithms (GA), several approaches
have been considered. The chapter presents a detailed discussion on available models
for Genetic-Fuzzy hybridization and applications developed based on existing
approaches. The extensive review on the genetic rule learning is presented as a part of
evolutionary fuzzy modeling. The evolutionary fuzzy modeling focuses on four major
approaches of Genetic- Fuzzy model, i.e. the Michigan, the Pittsburg, the Iterative
Rule Learning (IRL) and Genetic Cooperative-Competitive Learning (GCCL). The
characteristics of all these approaches have been discussed. The chapter
accommodates a discussion on the sub components of major hybrid approaches along
with their limitations and comparisons among one another. In the area of intelligent
decision support system, major application domains are taking advantages of machine
learning methods, especially Genetic-Fuzzy hybridization. A survey during research
in the area of various designed applications is documented. This investigation covers
several important applications domains where the machine intelligence is required to
be built. The real life applications of varied domains such as classification, medicine,
control systems, robotics, travel industry, stock and share, networking, etc employ
hybrid structures of GFS in order to achieve optimized rule learning. The chapter
finally draws conclusions from the current state of the art in the application of rule
learning from Genetic-Fuzzy hybridization. Subject to the conclusion made, the
chapter justifies the scope of the research carried out and reported in this thesis. As a
result of extensive research review on Genetic-Fuzzy hybridization, it has been
observed that no generalized framework using evolutionary fuzzy approach has been
developed in the field of education to solve the problems which lack mathematical
formulation.
3.2 Importance of Soft Computing
Soft Computing (SC) is not merely a clearly defined field but also a discipline that
deals with hybrid intelligent systems [174, p.239]. SC techniques are integrated
techniques to find solutions for the problems which are highly complex, ill defined
and difficult to model. It has been found that real world problems carry rapidly
changing information as well as they are somehow imprecise and uncertain in nature.
Chapter 3: Evolutionary Fuzzy Modeling
75
A Genetic-Fuzzy Approach to Measure Multiple Intelligence
To provide solutions for such kind of tasks using hard computing or traditional
computing is considered to be highly expensive and improper. It has been seen that
even after utilization of hard computing paradigm; real life applications suffer from
infeasible, non traceable and costly solutions to achieve high precision in their
outcome. This becomes the prime reason behind the transition from hard computing
to soft computing. Soft computing techniques are meant to operate in an environment
that is subjected to uncertainty and imprecision. According to Zadeh [126, p.1], the
guiding principle of soft computing is:
“Exploit the tolerance for imprecision, uncertainty, partial truth, and approximation to
achieve tractability, robustness, low solution cost and better rapport with reality.”
The family of soft computing is constructed using four prime techniques: namely
Fuzzy Logic (FL), Evolutionary Computation (EC), Neural Networks (NN) and
Probabilistic Reasoning (PR) as stated earlier in Chapter 1 (refer Figure 1.1 of
Chapter 1). Due to the characteristics of above mentioned techniques; soft computing
is distinguished from hard computing. These constituents are more flexible and robust
in providing solutions for real life problems. All four methodologies EC, FL, NN and
PR constitute the empire of soft computing which have been conceptualized and
developed before past fifty years. Each method is capable of providing distinguished
as well as sharable advantages and obviously carries certain weaknesses also. They
are considered complementary rather than competitive as desirable features lacking in
one approach are present in another. Initially, they had been applied for complex tasks
in isolation [166, p. 80]. As stated in earlier chapters, Evolutionary Computation (EC)
refers to the computer-based problem solving systems that use computational models
of evolutionary process. In order to implement evolutionary processes, EC includes
four methods namely Genetic Algorithms (GA), Evolutionary Strategies (ES),
Evolutionary Programming (EP) and Genetic Programming (GP). GA is one of the
prime optimization techniques from the tree of evolutionary search and optimization
that performs parallel, stochastic but direct search method to evolve the best solution.
Table 3.1 presents advantages and limitations of the principle constituents of Soft
computing that does provide a significant reason for inclusion of the constituents
[143, p.261 ; 63, p.390].
Chapter 3: Evolutionary Fuzzy Modeling
76
A Genetic-Fuzzy Approach to Measure Multiple Intelligence
Table 3.1: Advantages and Limitations of Principle Constituents of SC
Constituents of
SC Advantages Limitations
GA Natural evolution and optimization Inability of storing and
handling imprecision
FL Approximate reasoning,
imprecision Inability of learning
NN Learning and implicit knowledge
representation
Inability for
optimization
PR Uncertainty Inability of learning
GA provides a means to encode and to evolve rule antecedent aggregation operators,
different rule semantics, rule- based aggregation operators and de-fuzzification
methods. Therefore, GAs remain today as one of the fewest knowledge acquisition
schemes available to design and in some sense, optimize Fuzzy Rule Based Systems
(FRBSs) with respect to the design decisions, allowing decision makers to decide
what components are fixed and which ones evolve according to the performance
measures [54]. Such characteristics lead to the design of an intelligent system with
GA fuzzy hybridization which is a promising research field of modern computational
intelligence concerned with the development of the next generation of intelligent
systems.
3.2.1 Hybridization in Soft Computing
Actually, soft computing constituents are derived from varied domains as logic,
biology, physiology and statistics. In the initial stage of development of soft
computing, independent methods were utilized among community of practitioners.
Their success progressively attracted researchers in the other fields also. Indeed, they
model in different extents natural processes such as evolution, learning, or reasoning
[27, p.31].
The integration of different learning and adaptation techniques is required to
overcome limitation of individual methods so that synergetic effects through
hybridization or fusion of these techniques can be made achievable. Due to the lack of
Chapter 3: Evolutionary Fuzzy Modeling
77
A Genetic-Fuzzy Approach to Measure Multiple Intelligence
common framework; most of these approaches need to follow an adhoc and domain
specific design methodology. However, they are observed to be very successful but
still it becomes difficult to compare such various types of hybrid systems and
evaluation of their performance [31].
3.2.2 Integration of Evolutionary Algorithms and Fuzzy Logic
Recent years have contributed to large number of new hybrid evolutionary systems.
There are several ways to hybridize a conventional evolutionary algorithm for solving
optimization problems.
Evolutionary computing is based on Evolutionary Algorithms (EA). Genetic
Algorithms being one of the prominent types of EA were not specifically designed as
machine learning techniques like other approaches such as neural networks but have
been successfully applied to many search, combinatorial and optimization problems.
However, it is well known that a learning task can be modeled as an optimization
problem, and thus can be solved through evolution which is efficiently offered by EA
[55]. However, one of the significant limitations of GA; shown in the Table 3.1 is an
inability to deal with imprecision. This limitation can be solved using integration of
GA with FL. The principles and operations of Genetic Algorithm and fuzzy
computations have been described in Chapter 1 and Chapter 2 of the thesis.
Fuzzy Logic provides a mathematical way to represent and deals with vagueness of
everyday life. FL is presented not as a control methodology, but as a way of
processing data by allowing partial set membership rather than a crisp set membership
or non-membership [22].With the help of Fuzzy Logic based system; the knowledge
representation has been possible in a human understandable way using linguistic rules
to explain decision processes; but at the same time, fuzzy systems are suffering from
inability of self learning [166, p.89] as well as requires documentation of knowledge
which needs further continuous maintenance. At the other end; Genetic Algorithms
provide robust search capabilities both global and local in complex spaces whereas
Fuzzy Systems present flexible inference methods in order to deal with imprecision
and uncertainty [117]. The linguistic representation of knowledge allows a human to
interact with an FS in an intuitive, seamless manner. Hence, hybridization of FL with
GA becomes essential to achieve advantages of both the aforementioned approaches.
Chapter 3: Evolutionary Fuzzy Modeling
78
A Genetic-Fuzzy Approach to Measure Multiple Intelligence
In this aspect of hybridization of GA and FL, two main hybrid approaches are
analyzed [166, pp. 86-88; 45, p.180] which are narrated as under:
1. Fuzzy Logic Assisted Evolutionary Algorithm
This approach is also known as Genetic Algorithms controlled by Fuzzy Logic. In this
approach GA utilizes FL in order to adjust its parameters or operators to improve its
performance. Inherent parameters of GA such as fitness function and stopping
criterion are fuzzified. Hence, computational resources can be optimized from over
usages. This approach is further classified into two popular and successful subtypes
which are narrated as under:
Adaptive GA that adapt control parameters
In this approach, FLCs are used to dynamically compute optimal values for the GA
parameters. The objective is to adapt GA optimization process [166, p.87]. The
example of this approach is explained as under:
The control parameters of GA such as mutation and crossover rates, population size,
etc. can be dynamically computed by Fuzzy Logic controller (FLC). Eg. H. Y. Xu and
G. Yukovich, have categorized the size of populations as small, medium and large
[90].
GA with fuzzified genetic operators
In this approach, different fuzzified versions of genetic operators have been proposed
such as fuzzy connective crossover for real-coded Genetic Algorithms [56], [60], [57]
and soft operators [85].
2. Genetic Fuzzy Systems
In this type of approach, GA is used to improve the performance of FLC. As GA is
computationally expensive, GA based tuning is generally carried out off-line but if
integration of FLC is obtained with optimization, it can be controlled online [45,
p.180]. In order to hybrid GA with FS, it is required to incorporate fuzzy knowledge
in GA. The research design focuses on genetic rule learning for fuzzified knowledge
representation. In order to achieve the same, available models of Genetic Fuzzy
systems are discussed in detail in the upcoming sections.
Chapter 3: Evolutionary Fuzzy Modeling
79
A Genetic-Fuzzy Approach to Measure Multiple Intelligence
3.2.3 Genetic Fuzzy Systems
A GFS is basically a fuzzy system augmented by a learning process based on
evolutionary computation, which includes any method of EC family such as Genetic
Algorithms, genetic programming, and evolutionary strategies [5].
The most extensive GFS type is the Genetic Fuzzy Rule Based System (GFRBS),
where GA is employed to learn or tune (optimizing parameter) different components
of a Fuzzy Rule Based System (FRBS). In the design of GFS, a GA is used to look up
the performance of Fuzzy Logic controller (FLC) but the performance of FLC
depends on its Knowledgebase (KB) consisting Database (DB) and Rulebase (RB)
[45, p.170]. In order to achive design of FRBS, tasks such as designing inference
mechanism as well as generation of fuzzy rule set (KB or FRB) are required to be
satisfied [166, p.89].
As referred in Table 3.1 FRBSs are not able to learn themselves, but require the KB to
be derived from expert knowledge. In order to remove such limitation, evolutionary
learning process becomes essential to employ to automate FRBS design. By utilizing
this type of learning process FRBS can be defined automatically. The stated type of
design can be considered as an optimization or search problem. In order to solve
optimization problems, GAs are selected due to major capabilities such as:
Being global search method, GAs can explore large a search space;
Able to find near optimal solutions in complex search spaces; and
Able to provide generic code structure and independent performance.
Due to the above mentioned capabilities, it is possible to incorporate a priori
knowledge in GA which may be in form of linguistic variable, fuzzy membership
function parameters, fuzzy rules, etc. Finally, Genetic- Fuzzy system can be designed
as shown in Figure 3.1.
Figure 3.1 illustrates the general structure of Genetic Fuzzy Rule Based System. The
design of GFRBS is constituted using three layers which are explained as under.
1. Interface Layer
The bottom layer of the Genetic-Fuzzy Rule Based System is named as interface
layer. It is composed of three major components: Environment, FRBS and Output
Interface. This layer is basically structured using input interface, comparison with
Chapter 3: Evolutionary Fuzzy Modeling
80
A Genetic-Fuzzy Approach to Measure Multiple Intelligence
fuzzy system and processed output of FRBS. Here, the input interface defines the
members of an application domain and interact with FL layer which is responsible for
designing and implementing the fuzzy system. The FL layer is explained as under.
2. FL Layer
The middle layer of the Genetic Fuzzy Rule Based System is named as FL layer. It
interacts with the interface layer to have input variables. This layer consists of various
components such as Fuzzification Interface, Inference Mechanism and
Defuzzification Interface. This layer is responsible to design the processes which are
related to fuzzy system implementation. The input interface decides the members or
the variables of the application domain in order to generate fuzzification. The
inference mechanism is designed using the members which are selected in the input
interface and will be involved in the process of generating the FRBS. The output
interface is responsible for defuzzification of input variables. It provides the results
produced by Fuzzy System (FS).
Figure 3.1: General Structure of Genetic Fuzzy Rule Based System
Environment Environment Comparison with Fuzzy Rule
Based Systems
Design Process
Genetic Based Learning Process
Knowledge Base= Rule Base + Data Base
Inte
rfac
e L
ayer
F
L L
a y e
r
Rep
osi
tory
Lay
er
o
Members or
Variables of
Application Domain
Input
Interface
Components
of Fuzzy
System
FRBS Output of
Processed FRBS
Output
Interface
Chapter 3: Evolutionary Fuzzy Modeling
81
A Genetic-Fuzzy Approach to Measure Multiple Intelligence
3. Repository Layer
The top most layer of the Genetic Fuzzy Rule Based System is named as repository
layer. This layer is responsible for designing of GFRBS. In order to design, GFRBS,
evolutionary techniques are required to be incorporated in order to achieve automatic
generation or modification of entire part of the Knowledge Base (KB). KB is a
combination of Data Base (DB) and Rule Base (RB). The parameters of
knowledgebase include fuzzy rules and membership functions. Both the constituents
interact with inference mechanism of the middle layer. In order to achieve
optimization, it is required to find an appropriate KB.
3.2.3.1 Genetic Learning Process
In the aspect represented by the Figure 3.1, designing GFS requires to incorporate
genetic learning process for designing or optimizing the knowledgebase. GFRBS is a
design method for FRBSs which incorporates evolutionary techniques to achieve the
automatic generation or modification of the entire or part of the knowledgebase. In
order to evolve rules, GA needs to search the genotype space as well as it also
requires some mechanism generating new variants from the currently existing
candidate solutions. The objective of the search process is to maximize or minimize a
fitness function that describes the desired effectiveness of the system.
In summary, the genetic process is the result of the interaction between the evaluation,
selection and creation of genetically encoded candidate solutions, which represents
the contents of the KB of an FRBS.
Approaches for Rule Base Learning with Genetic Algorithm
Generating system which employs learning approaches is capable enough of changing
their underlying structure with the objective of improving their performance or quality
of their knowledge according to certain criteria. With reference to genetic rule
learning using GA in order to optimize FLC, major approaches are identified as
follows:
1. Approach A: GA based optimization for automatic constructed FLC
The approach “A” is mainly suitable for solving complex tasks. For a complex task, it
is very difficult to design KB manually because in such cases performance of GFS
cannot be achieved as per expectations. Automatic design of FLC through GA
Chapter 3: Evolutionary Fuzzy Modeling
82
A Genetic-Fuzzy Approach to Measure Multiple Intelligence
becomes solution for complex tasks. GA utilizes information of DB, RB and
consequent part of each rule of the FLC. The predefined training set is utilized by GA
strings to optimize FLC. As a result, optimal FLC can be evolved through iterations.
To design system based on such approach is a critical task.
2. Approach B: GA based optimization of manually constructed FLC
The working characteristic of the approach “B” is based on designer‟s knowledge of
the process to be controlled. In order to process linguistic knowledge representation,
ranges of variables are decided and different linguistic terms are designed as per
requirements of the problem. The number of input combinations for the design of
FLC depends on the number of input variables and their linguistic terms. In order to
propose the DB of the FLC manually, it is needed to have distribution of membership
functions and variables by designer and appropriate design of RB is made possible,
later. This type of design of FLC may be flexible but it is not optimal as well [45,
p.174]. In this approach, a GA is used to tune the DB and/or RB of the FLC with the
help of training scenarios. After the GA-based tuning is over, the FLC will be able to
determine the output for a set of inputs within a reasonably accuracy limit.
Figure 3.2 represents parallel approaches namely “A” and “B” along with their
characteristics for developing Genetic- Fuzzy systems.
Figure 3.2: Approaches for GA based Optimization for FLC
Approach A: GA Based Optimization for Automatically
Constructed FLC
• Handles complex tasks
• Utlizes knowledge of experts of GA
• More complex
• Must generate an optimal FLC
• More time consuming process from execution of GA perspective
Approach B: GA Based Optimization for Manually Constructed FLC
• Handles easy task
• Utilzes knowledge of designer
• Less complex
• May not generate an optimal FLC
• Less time consuming process from execution of GA perspective
Chapter 3: Evolutionary Fuzzy Modeling
83
A Genetic-Fuzzy Approach to Measure Multiple Intelligence
3.3 Techniques of Hybridization for GFS
Although, GA was not specifically designed for learning, they are used as global
search algorithms. Apart from the searching task, they do offer a set of advantages for
machine learning. Many methodologies for machine learning are based on the search
of a good model and they are very flexible because the same GA can be used with
different representations [54, p.33]. In a rule based system, to achieve task of learning
rules, there are two major styles available to encode rules within populations of
individuals of GA as shown in Figure 3.3.
Style 1: The ‘‘Chromosome = Rule’’ Approach
In the style 1, each individual codifies a single rule, and the whole rule set is provided
by combining several individuals in a population known as rule cooperation or via
different evolutionary runs known as rule competition [101]. For example, the
Michigan approach and the Iterative Rule Learning (IRL) approach are representative
approaches of Style 1.
Figure 3.3: Popular Styles of Rule Encoding in Genetic-Fuzzy Hybrid Systems
Style 2: The „„Chromosome = Set of Rules‟‟ Approach
In style 2, each individual represents a rule set so it is popularly known as the
“Chromosome = Set of rules”. In this case, a chromosome evolves a complete RB and
they compete among them along with the evolutionary process. E.g. the Pittsburgh
approach is the representative approach of style 2 [183].
Style 1
Chromosome="Single Rule"
Ex. The Michingan Approach
Ex. The IRL Appraoch
Style 2
Chromosome= "Set of Rules"
Ex. The Pittusburg Approach
Chapter 3: Evolutionary Fuzzy Modeling
84
A Genetic-Fuzzy Approach to Measure Multiple Intelligence
In order to provide rule base learning process with different levels of complexities,
evolutionary algorithm provides three major approaches for evolving such rule
systems namely the Michigan Approach, the Pittsburg Approach, the Iterative Rule
Learning (IRL) approach as shown in Figure 3.4.
Figure 3.4: Major Approaches for Genetic Fuzzy System
3.3.1 The Michigan Approach - Classifier System
Learning Classifier System (LCS) is a machine learning technique which combines
reinforcement learning, evolutionary computing and other heuristics to produce
adaptive systems. This approach represents “Chromosome=rule” approach as shown
in Figure 3.3. In order to generate populations of rules automatically, Classifier
Systems (CSs) or production rule systems are designed [99, pp. 263-296]. Learning
Classifier Systems are a kind of rule-based system with general mechanisms for
processing the rules in parallel, for adaptive generation of new rules, and for testing
the effectiveness of existing rules [100]. These mechanisms make possible
performance and learning without the “fragile” characteristics of the most expert
systems in AI [102, p. 3].
CSs are especially parallel, message passing, rule based systems that learn through
credit assignment and rule discovery [166, p.130]. CSs typically operate in
environments that exhibit one or more of the following characteristics:
The Michigan Approach
The Pittusburg Approach
The Itreative Rule Learning (IRL)
Approach
Genetic Fuzzy System
Chapter 3: Evolutionary Fuzzy Modeling
85
A Genetic-Fuzzy Approach to Measure Multiple Intelligence
Large amount of irrelevant or noisy data can be incorporated;
Uninterrupted real time required for action; and
Implicit or vague goals.
Figure 3.5 shows comparison of traditional expert system and classifier system along
with characteristics of each of them. These characteristics are narrated as under.
CSs are different than traditional expert systems by providing several advantages such
as rules are needed to design by knowledge engineer in expert system while in
Classifier System (CS), classifier rules are generated by a GA. In the expert system,
rules are stored and processed in a simple way as well as no additional parameter for
rule system are available to calculate rule strength as well as evolution of it cannot be
achieved. Each classifier has to pass through performance evaluation system and have
a strength parameter; a number is assigned to each classifier, after that. Figure 3.5
distinguishes characteristics of expert system and classifier system.
Figure 3.5: Comparison of Traditional Expert System and Classifier System
The prototype organization of Classifier Systems (CSs) is composed of three sub
systems [194, p.49] which are narrated as shown below:
Traditional Expert System
(ES)
• Rules are normally designed by knowledge engineer
• Rules have condition and action parts
• Decision making is possible
• No evolution is achieved
Classfier System
(CS)
• Classfiers are generated through GA
• Message list is associated with each of the classifiers
• Message list contains current state and internal message
• Each classifer has strength associated in form of number
Chapter 3: Evolutionary Fuzzy Modeling
86
A Genetic-Fuzzy Approach to Measure Multiple Intelligence
1. A production system with a rule base which processes incoming messages from
the environment and sends output messages to the environment.
2. A design of credit system which receives pay-off from the environment and
determines which rules had been responsible for the feedback.
3. A Genetic Algorithm which recombines existing rules and introduces new rules.
Figure 3.6 illustrated above mentioned subsystems along with major task of each one.
Figure 3.6: Subsystems of Classifier Systems (CS)
After implementing the above sub-systems, the working of classifier system is
summarized as below:
The working of classifiers is strengthened through the Credit Assignment (CA)
system. Basically, a GA selects high fitness classifiers as parents forming offspring by
recombining components from the parent classifiers. Here, traditional fitness function
of GA is not used; instead the fitness of a classifier is determined by strength
calculated with the CA system. In typical CS implementations, from the set of
classifiers, high strength classifiers constitute the GA population. The strategy of GA
is to replace the worst set of classifiers by newly created strong classifiers. Due to this
strategy, the high performance of the classifier can be achieved [166, pp.132-136].
The following approaches are the fundamental approaches based on evolutionary
fuzzy modeling using the Michigan approach.
Evolutionary fuzzy modeling has since been applied to an ever-growing number of
domains, branching into areas as diverse as chemistry, medicine, telecommunications,
biology, and geophysics. Some applications are briefly narrated as follows:
The Production System
(Rule Base)
The Credit Assignment System
(Assignment of Strength to Classifier )
The Classifier Discovery System
(Learning of Classifiers through
GA)
Chapter 3: Evolutionary Fuzzy Modeling
87
A Genetic-Fuzzy Approach to Measure Multiple Intelligence
A Genetic Algorithm based learning process has been proposed to determine the
optimal data base of a Fuzzy Logic controller keeping its rule base which is
previously defined [34].
In 1999, H.Ishibushi,et al.[79] have presented fuzzy classifier systems for high
dimensional pattern classification problems and in 2000, Juang et al. have presented
genetic learning for fuzzy controller design [30]. In 1999, Y.Shi et al. discussed GA
based method to evolve fuzzy expert system for large complex problems. This system
can evolve the rule set along with optimal number of rules inside it, tune the
membership functions and evolve the membership function [203].
M. Valenzuela- Rendon , has proposed linguistic RB learning using Michigan
approach [148]. This proposal presents the first GFS based on the Michigan approach
for learning RBs with DNF fuzzy rules. This approach incorporates a reward
distribution scheme that requires knowledge of the correct action, and thus, must be
considered as a supervised learning algorithm [149].
In GFS, the fuzzy inference system is represented by the entire population having
several rules participating. In order to propose the best action, these rules are in
constant competition and cooperate to form an efficient fuzzy system. Identification
of the specific rules responsible for good system behavior is a very difficult task. This
task requires design of fitness function capable of measuring goodness of a single
fuzzy rule as well as the quality of its cooperation with other fuzzy rules in the
population to give the best action as an output [27, p.33]. The design of a fitness
function of this kind is not an easy task [4]. In 1978, the learning classifier systems
based on Michigan approach are designed by J. H. Holland & J. S. Reitman [101]. In
1988, K.A. De Jong has analyzed two different approaches for learning the rule set of
production system [110]. In 1991, the significant research work on the classifier
systems has been proposed by M.Valenzuela- Rendon [148].
Optimum design of Fuzzy Logic controllers using Genetic Algorithms is presented by
[48]. Here, the GA is used to modify the fuzzy relational matrix of a one-input, one-
output fuzzy model using fuzzy relation matrix rather than the decision table. The
major limitations of The Michigan approach are presented as follows:
Chapter 3: Evolutionary Fuzzy Modeling
88
A Genetic-Fuzzy Approach to Measure Multiple Intelligence
Limitations of The Michigan Approach
Michigan approach is mainly concerned with continuous (on-line) learning in
non inductive problems. One of the major limitations is the risk involved due
to random modification done by GA for applications especially designed to
achieve security [194, pp.54-55]. The reason behind this limitation is
explained as under:
Online rule learning process is achieved in which every single rule is
manipulated and hence requires an intense analysis of the performance of
every rule. Such a system can adapt to varying environmental circumstances
automatically. In this case, the rule is modified every time.
The other limitation observed is: the Michigan approach based system may
fail if working with complex environment; i.e. there is only a low probability
that important state sequences are observed repeatedly.
Further, this approach represents the knowledge of a single entity that learns
through interaction with the environment and later being adapted to it rather
than to evolution of possible solutions in form of rules [166, p.142].
3.3.2 The Pittsburgh Approach
It represents “Chromosome is Rule set” approach as shown in Figure 3.3. The central
idea is to impart intelligence through evolution. In this approach, an evolution can be
generated through competition among the individuals and adaptation to the
environment [166, p.142].
This approach is particularly tailored for training in both inductive and non inductive
problems. In the Pittsburgh approach, each individual represents a complete entity of
knowledge and due to this type of structure; different individuals do not require
interacting with one another for the evaluation of the knowledge. Hence, the need of
credit assignment, and, thus the definition of complex algorithms with that purpose
are not required. The individuals responsible in generating the evaluation are assigned
the performance measure. Figure 3.7 represents the block diagram of subsystems of
Pittsburg rule learning approach.
Chapter 3: Evolutionary Fuzzy Modeling
89
A Genetic-Fuzzy Approach to Measure Multiple Intelligence
Figure 3.7: The Subsystems of the Pittsburg Rule Learning Approach
The major components of Figure 3.7 are described as follows:
Rule base and the Population of RBs
In order to generate the learning process, it is required to generate a population of
potential solutions to the problem. The population of potential solutions is designed
from RBs through a common processing structure to solve a specific problem. Each
RB in the population is evaluated by applying it to solve the problem. Feedback is
generated from the environment. Each RB is evaluated independently and no
interaction between individuals of the population occurs during the evaluation. To
start the learning process, an initial population of RB is required. In some cases, it is
obtained from available knowledge, while in other cases; it is randomly generated
[166, pp.143-144].
The Evaluation System
As stated in the Pittsburg system, each RB is to be evaluated. This evaluation of RB is
based on the effect of the interaction of the rule based system, applying the
corresponding RB, with the environment. As a result of this interaction, the
environment generates a feedback that is used by the evaluation system to generate
the evaluation of the RB. The evaluation is quite different depending on the
application and the environment [166, p.144]. The evaluation system often becomes
the most time consuming element of the process.
Comparatively, this approach offers more simplicity compared to the Michigan
approach by producing independent evaluation of population. Here, larger
Rulebase System
Population of RB
RBs Discovery
System
Evaluation System
EnvironmentFeedback
Chapter 3: Evolutionary Fuzzy Modeling
90
A Genetic-Fuzzy Approach to Measure Multiple Intelligence
computational efforts for independent evaluations are required; hence this approach is
free from conflicts, generated as a result of interaction.
The Rule Base Discovery System
Once the process of evaluation of population of RB is completed, new RBs are to be
searched, and, RB discovery system is to be introduced. This system generates a new
population along with set of genetic operator to the previous generation [166, p.144].
The rule base discovery system for the Pittsburg approach is different from the
Michigan approach by the following characteristics:
The Level of Replacement
In the Michigan approach, the number of replaced individuals at each generation has
to be low enough to preserve the system performance, as it is the result of the
interaction between the individuals whereas in Pittsburgh approach the performance
of the best individual is achieved and consequently due to that performance, there can
remain steady as long as the best individual is maintained.
The Timing of Evaluation and Discovery
In the Michigan approach, discovery is applied with a lower frequency than credit
assignment. In this approach, continuous learning is made possible to reach a steady
state situation before creating a new generation while the Pittsburg approach
implements predefined training cycle which has to be applied for each individual in
the population. Consequently, the discovery phase takes place after a complete
training cycle for each individual [166, pp.144-145].
The Pittsburg approach facilitates to include additional optimization criteria in the
fitness function, thus affords the implementation of multi-objective optimization. In
1991, Thrift [172] has presented a pioneer work on the Pittsburgh approach for
learning Rule Base. This method works by using a complete decision table that
represents a special case of crisp relation defined over the collections of fuzzy sets
corresponding to the input and output variables. Finally, the GA employs an integer
coding.
Summarizing the above mentioned characteristics of the Michigan and the Pittsburg
approaches, Figure 3.8 is constructed which shows distinguishable characteristics of
the Michigan approach and the Pittsburg approach:
Chapter 3: Evolutionary Fuzzy Modeling
91
A Genetic-Fuzzy Approach to Measure Multiple Intelligence
Figure 3.8: Comparison of the Michigan and the Pittsburg Approaches
Different kinds of Rule based systems have been designed and evolved using this
approach among them significant examples are presented as follows:
Concept learning can be evolved by GABIL system which consists of RBs
with boolean rules [111]. This system continuously learns and refines concept
classification rules form its interaction with environment.
Genetic Based Inductive Learning (GIL) systems designed by C.Z.Janikow,
utilize coding of RBs using logical representation for learning concept
descriptions [38].
Classification with real valued representation for continuous variables that
utilizes RBs using Genetic Algorithm is achieved [15].
Evolutionary learning of fuzzy rules (ELF) method is proposed and applied to
the problem of guiding an autonomous robot by A. Bonarini in 1993. The key
The Michigan Approach
It operates on single rule in an
online process or a simulated
environment
It concerns with online learning in
non inductive problems
Candidate solution is embeded in
performance system
The whole population consistutes
RB and single entity is evaluated
through performance system
Complexity of fitness evaluation
is high
The PittusBurg Approach
It operates on multiple rules constitued in FRBS
It concerns with training of
inductive and non inductive
problems
Candidate solutions exist in a
seperate entity
Entire population of RB is
evaluated once at a time &
feedback is assigned to RB under
evaluation
Complexity of fitness evaluation
is low
Chapter 3: Evolutionary Fuzzy Modeling
92
A Genetic-Fuzzy Approach to Measure Multiple Intelligence
issue of ELF is to find a small rule base which only contains important rules
[4].
Limitations of The Pittsburg Approach
The Pittsburg approach evaluates the population of the entire fuzzy system for every
generation. It evaluates fitness for each RB of the population and repeats the same
process for the entire RB, thus demanding more computational resources. Hence, the
computational cost of this approach is high, which is the major limitation of the
Pittsburg approach [27, p.34].
3.3.3 The Iterative Rule Learning Approach (IRL)
The major drawback of the Michigan approach and the Pittsburg approach is the
consumption of huge amount of computer memory for searching numerous fuzzy
rules. To overcome the above stated problem, an iterative rule learning approach
(IRL) is designed. The IRL (Iterative Rule Learning) approach is based on the
approach of “Chromosome is rule”. The prime reason to develop this approach is to
integrate the best features of the Michigan and the Pittsburg approaches. In this
approach, a new rule is added to the rule set, in an iterative fashion, for every run of
GA. The Itreative Rule Learning (IRL) approach was first proposed by Supervised
Inductive Algorithm (SIA) designed by G.Venturini [68] in 1993 and it has been
widely developed in the field of GFRBs [8, 13,162,165].
This approach works by combining the styles of the Michigan and the Pittsburg
approaches. Similar to the Michigan approach, each chromosome in the population
represents a single rule, but similar to the Pittsburg approach, only the best individual
are considered to form part of the final solution. As a result of this type of
computation, the generated RB ultimately discards the remaining chromosomes in the
population. Therefore, in the iterative model, the GA provides a partial solution to the
problem of learning, and, it is repeated multiple times to obtain the complete set of
rules. Hence, the chances of generating best solution are possible [166, pp.148-149].
3.3.3.1 Multistage Processes of Iterative Rule Learning
The learning process utilizing Iterative Rule Learning (IRL) is composed of two
major stages. This type of GFRBSs based on IRL approach is popularly known as
Chapter 3: Evolutionary Fuzzy Modeling
93
A Genetic-Fuzzy Approach to Measure Multiple Intelligence
multi-stage GFRBSs [8]. Figure 3.9 represents two pioneer stages required to develop
IRL.
Figure 3.9: Different Stages of IRL
In order to generate an RB, which constitutes a true solution to the learning problem,
the GA is embedded into an iterative scheme similar to the following [194, pp.43-44]:
Step 1: Use a GA to obtain a rule for the system.
Step 2: Incorporate this rule into the final set of rules.
Step 3: Penalize this rule.
Step 4: If the set of rules generated till now is adequate to solve the problem,
end up returning it as the solution. Otherwise, return to step 1.
An evolutionary algorithm is used to find a single rule, thus providing a partial
solution. The evolutionary algorithm is then used iteratively for the discovery of new
rules, until an appropriate rule base is built. A penalization process is applied each
time a new rule is added. As an effect of this process, redundant rules are avoided [27,
p.34]. Chromosomes compete with every GA run, choosing the best rule per run. The
global solution is formed by these best rules. Learning algorithms that use the IRL
approach do not imagine any relationship between them in the process of obtaining
rules. Therefore, the final set of rules usually needs an a posteriori process which will
modify and/or fit the said set [45, pp. 33-34].
In order to implement a learning algorithm based on GAs using the IRL approach,
following sub components are required [166, p.150].
A Generation Process
• Representation of knowledge in
data set through chromosome
• Execution of Iterative covering
method
• Execution of Fuzzy rule
generating method frequently to
find best rules
A Post Processing Process
• Simplification of the rule set
• Selection and refinement of
rules
• Refinement of previous rule set
to remove redundent rules
• Demonstration of best
performance of final rule set
Chapter 3: Evolutionary Fuzzy Modeling
94
A Genetic-Fuzzy Approach to Measure Multiple Intelligence
Figure 3.10 represents sub components of IRL which are discussed as follows:
Figure 3.10: Sub Components of IRL
1. A criterion for selecting the best rule in each iteration
This component is used to determine the good rules. The selection criteria about the
rule strength are also presented which includes number of examples covered, criteria
of consistency of the rule or criteria of simplicity.
2. A Penalty Criterion
This component of criterion is often associated, although it is not necessary, with the
elimination of the examples covered by the previous rules.
3. A Determination Criterion
This component is used to determine the assurance about the completeness of the set
of rules. It starts working when enough rules are available to represent the examples
in the training set. The main functionality of this component is to check whether all
the examples in the training set are sufficiently covered or not.
Advantages offered by IRL Approach
The significant advantage of IRL is that it reduces the search space, because in
each sequence of iterations, the learning method only searches for a single best
rule instead of the whole RB.
This approach combines the speed of the Michigan approach with the
simplicity of the fitness evaluation of the Pittsburgh approach.
A Criterion for
Selecting the Best
Rule
A Penalty Criterion
Determination
Criterion for
Training Set
Chapter 3: Evolutionary Fuzzy Modeling
95
A Genetic-Fuzzy Approach to Measure Multiple Intelligence
The Michigan approach provides online learning for non-inductive learning
problems whereas IRL approach facilitates off line inductive learning
problems.
There are different examples of IRL approach for genetic rule learning in the
literature; the major among them are presented below:
Supervised Inductive Algorithm (SIA) uses a single GA that goes on detecting the
rules and eliminating the examples covered by the latter. SIA can only work with
crisp data [68]. MOGUL (Methodology to obtain GFRBSs under the IRL approach)
has been proposed for designing different types of FRBS for variety of the problem
domains i.e fuzzy modeling, fuzzy control and fuzzy classification [61,158, 164]. In
order to work with MOGUL, a user has to define evolutionary process in each of the
GFRBS learning stages. MOGUL works with different fuzzy models i.e. descriptive
Mamdani type [158, 163] and approximate Mamdani type [61, 157] as well as TSK
FRBSs [159].
SLAVE (Structural Learning Algorithm in Vague Environment) is a genetic learning
process based on IRL approach to design DNF Mamdani type FRBSs that was
proposed by A.Gonza'lez in 1993 [13] and later refined in various research papers of
A.Gonza'lez, and R.Perez [9,10,11,12,128]. SLAVE launches a new GA to find a new
rule after having eliminated the examples covered by the last rule obtained. SLAVE
was designed to work with or without linguistic information.
The genetic generation process runs a GA for obtaining the best rule according to
different features, assigns a relative covering value to every example, and removes the
examples with a covering value greater than a constant [58,59].
3.3.4 Other Approaches of Genetic Rule Learning
Another major approach is identified as Genetic Cooperative-Competitive Learning
(GCCL) to provide genetic rule learning. In this approach, the rule base can be
encoded with the help of a complete population or a subset of population [54, p.34].
In this model the chromosomes compete and cooperate simultaneously with one
another.
Chapter 3: Evolutionary Fuzzy Modeling
96
A Genetic-Fuzzy Approach to Measure Multiple Intelligence
COGIN (Coverage-based Genetic Induction) proposed by Greene and Smith in
1993[47], REGAL [7] and LOGENPRO [142, p.71] are the successful examples of
GCCL kind of learning approach.
3.4 Work done so far in the Area of GFS Hybridization
It has been observed that there are many real world applications successfully designed
to support the need of intelligent decision support and to achieve optimization using
Genetic-Fuzzy hybridization. The major applications obtained as a result of the above
mentioned approaches and explained as follows:
Pattern Recognition constitutes an important application of FRBS as FRBSs provide a
suitable mean to classify incomplete and imprecise data [129, pp.15-36; 204, pp.152-
178].
As mentioned earlier, SLAVE has been applied to solve classification problem of
myocardial infarction [14]. The set of the linguistic rules generated by SLAVE is
more understandable and more useful to the clinical experts. SLAVE has also been
applied to diagnose malignancy. In order to diagnose breast cancer in females,
Wisconsin Breast Cancer Diagnosis (WBCD) database is designed [33] using the nine
characteristics with 683 instances representing the classification of malignant
attributes. The diagnostic accurate classifier system using evolutionary fuzzy rules has
been developed by C. A. Pena-Reyes and M. Sipper [28].
C.H.Wang,et al.proposed several GA-based knowledge integration strategies to
automatically integrate multiple rule sets in a distributed-knowledge environment.
Also, a self-integrating knowledge-based brain tumor diagnostic system based on
these strategies was successfully developed. Here, Genetic Algorithm generates an
optimal or nearly optimal rule set from these initial knowledge inputs. Furthermore, a
rule-refinement scheme is proposed to refine inference rules via interaction with the
environment [32].
The Genetic-Fuzzy rule base system (GFRBS) using the Pittsburg approach has also
been designed for classification in breast cancer diagnosis [166, pp.377-378]. Here,
KB consists of membership function and its parameters. Later, it identifies the best
single rule for classification.
Chapter 3: Evolutionary Fuzzy Modeling
97
A Genetic-Fuzzy Approach to Measure Multiple Intelligence
GFS modeling has been used to estimate human dental age from tooth eruption status
and patient chronological ages. Here, an FRBS is compared with the expert
knowledge and is automatically generated from a set of data collected by the
researcher. GFRBS based on the Pittsburg Approach has been designed [136,137].
GFRBS for linguistic and fuzzy modeling has been achieved in different applications
for variety of domains, which are briefly narrated as follows:
There are two different genetic learning processes consist of two stages are designed
for linguistic FRBS based on first generating RB. These two RB learning processes
considered for the first stage are: WM adhoc data driven method and Thirft‟s GFRBS
[16,165,158]. An adaptation of the genetic learning process is designed [78] which
generates multiple RBs from different fuzzy partitions and joining them in a global
RB and applying a genetic selection process to obtain the best sub set of rules [167].
In 1994, the subjective qualification of rice tastes using sensory test is designed by H.
Ishibushi et al.. This test has been performed to identify quality of rice using genetic
fuzzy classification [77, 86].
In the field of behavior based robotics, GA based system has been designed for
robotic system in order to navigate, plan and operate in the real world with intelligent
behavior. Automatic generation of design of robotics system is possible through
Evolutionary approaches. In order to achieve automatic generation of design, the
objective function is designed such a way that it is capable to observe the robotic
behavior evolved by the means of the EA [166, pp.414-416].
A learning classifier system based on Michigan-style Learning Fuzzy-Classifier
System (LCS) has been designed for the supervised learning tasks. Here, the problem
of interpretability in LCSs, and design of Fuzzy-UCS - an online accuracy-based
LFCS architecture has been addressed [17].
In order to optimize social regulation policies, the decision support system is designed
by A. Abraham et al., which is popularly known as EvoPOL. EvoPOL is a fuzzy
inference based decision support system that uses an evolutionary algorithm (EA) to
optimize the if-then rules and its parameters for social regulation policies[2, pp.814-
826].
A computational intelligence framework has been designed to cope with the problem
Chapter 3: Evolutionary Fuzzy Modeling
98
A Genetic-Fuzzy Approach to Measure Multiple Intelligence
of controlling the dynamical model for a disease in order to provide some
understanding related to the medical treatment. A genetic fuzzy system approach to
control a nonlinear dynamic model of the HIV infection is presented by [139].
The web based decision support system popularly known as WBISC-DSS is an
admission process which is designed using evolutionary fuzzy computing for
university of Berkeley, California. This system has been developed to allow automatic
adjusting of the user‟s preferences. These preferences can be seen as the parameters of
the Fuzzy Logic model in the form of degrees of importance of the used variables
[182].
M. A. Lee & H. Takagi have proposed the design method for fuzzy system which uses
Genetic Algorithm. Here, GA integrates three design stages; membership functions,
the number of fuzzy rules, and the rule consequent parameters to obtain optimal fuzzy
system [137].
A decision model for trading system has been designed that combines Fuzzy Logic
and technical analysis to find patterns and trends in the financial indices [198]. The
fuzzy model is optimized by utilizing a Genetic Algorithm and the historical data.
The classification of rules in dermatology data sets and breast cancer data set for
medicine has been discovered by GA. In this work; skin diseases like psoriasis,
seboreic dermatitis, lichen planus, pityriasis rosea, chronic dermatitis and pityriasis
rubra pilaris are discovered by implementing several if-then rules using GA. This
system has also been implanted in order to determine the patients suffering from
breast cancer. The goal of this research work is to find out the patients to whom the
cancer may re-occur [147].
The bilingual question classification through Genetic Algorithm with machine
learning methods has been designed with an Integrated Genetic Algorithm (GA) and
Machine Learning (ML) approach for the question classification in cross-language
question answering [150].
An evolutionary system is designed for discovering fuzzy classification rules. In order
to perform classification task, three methods were used to extract fuzzy classification
rules using EA. These three methods are genetic selection of fuzzy rules from a large
number of fuzzy candidate rules, genetic reduction of genetic space, and genetic
learning of fuzzy classification rules [141].
Chapter 3: Evolutionary Fuzzy Modeling
99
A Genetic-Fuzzy Approach to Measure Multiple Intelligence
In order to predict, IPO under pricing, a prediction system has been designed using
the Genetic Algorithm. The Michigan approach has been utilized for generating the
Evolutionary Rule-Based System for IPO under-pricing [49].
A Genetic- Fuzzy hybridization approach has been implemented in travel choice
behavior models to provide insight into the travel choice behavior that is required in
design and planning of transport systems of the urban public transport networks [184].
An approach to a classical network optimization problem or the transportation
problem has been designed to optimize the distribution network for identifying
shortest route for the distribution of vehicles. In order to achieve optimization,
Genetic Algorithm is designed while fuzzy sets have been used to represents the
imprecise information related to provisional information such as costs, demands and
other variables [52].
3.5 Conclusion
The chapter presents the significance of hybridization of soft computing methods in
order to design solution for real life applications with desired characteristics. Being
two prime constituents, GA and FS provide advantages for machine learning, search
and optimization. Here, the need of hybridization of GA with FS is explained. In
order to achieve such integration, two major GA-Fuzzy hybrid approaches have been
analyzed: Fuzzy Logic assisted evolutionary algorithm and Genetic- Fuzzy Systems.
The chapter explains GFS in broad ranges of approaches for genetic rule learning. The
general structure of GFS has been presented along with two major approaches for
optimized rule learning using GA for automatic generated FLC as well as manually
designed FLC. The techniques of Genetic-Fuzzy hybridization are structured using
two popular styles of encoding:”Chromosome is rule “approach and “Chromosome is
set of rule “approach. Genetic-Fuzzy hybridization approaches are designed using
either of the two above stated styles. GFS hybridization is classified into three major
approaches: the Michigan, the Pittsburg and The IRL. The chapter explains each of
these approaches along with subcomponents, working style, advantages and
limitations. It also compares characteristics of traditional expert system with
Classifier System for Genetic- Fuzzy hybridization and hence presents the importance
of designing GFS. Comparative evaluation of each of the hybrid method leads to the
design of IRL. The chapter also narrates various advantages of the Iterative Rule
Chapter 3: Evolutionary Fuzzy Modeling
100
A Genetic-Fuzzy Approach to Measure Multiple Intelligence
Learning approach. Extensive literature survey of GFS on the varied domains such as
classification, medicine, machine learning, management, etc. has been presented.
In summary, it can be stated that hybridization of Genetic-Fuzzy approaches is
efficient in order to achieve rule learning in an optimized way and hence can be one
of the most suitable methods for the machine learning. In order to design a system,
based on genetic rule learning from fuzzified input, Genetic-Fuzzy System is
required. The approach similar to IRL is required to achieve the genetic rule learning
using FLC. It has been observed from the literature review on GFS, that real life
problems which are based on mathematical formula can be easily designed and
implemented using GA with FLC. The significant observation also says that there is
not any generic framework yet developed for problem lacking mathematical
formulation using genetic fuzzy hybrid approach order to achieve optimization for
rule learning The research work is a step towards the same.