Authors:
Roman Sergienko
1
and
Eugene Semenkin
2
Affiliations:
1
Ulm University and Siberian State Aerospace University, Germany
;
2
Siberian State Aerospace University, Russian Federation
Keyword(s):
Fuzzy Classifier, Michigan Method, Pittsburgh Method, Coevolutionary Algorithm, Self-tuning, Strategy Adaptation, Multistep Procedure.
Related
Ontology
Subjects/Areas/Topics:
Artificial Intelligence
;
Artificial Intelligence and Decision Support Systems
;
Computational Intelligence
;
Enterprise Information Systems
;
Evolutionary Computing
;
Genetic Algorithms
;
Hybrid Learning Systems
;
Informatics in Control, Automation and Robotics
;
Intelligent Control Systems and Optimization
;
Knowledge-Based Systems Applications
;
Machine Learning in Control Applications
;
Soft Computing
Abstract:
A method of Michigan and Pittsburgh approaches combining for fuzzy classifier design with evolutionary algorithms is presented. Michigan-style stage provides fast search of fuzzy rules with the best grade of certainty values for different classes and smoothing of randomness at initial population forming. Pittsburgh method provides rules subset search with the best performance and predefined number of the rules and doesn’t require a lot of computational power. Besides self-tuning cooperative-competitive coevolutionary algorithm for strategy adaptation is used on Michigan and Pittsburgh stages of fuzzy classifier design. This algorithm solves the problem of genetic algorithm parameters setting automatically. The next result is multistep fuzzy classifier design based on multiple repetition of previous fuzzy classifier design. After each iteration standard deviation of classification performance decreases and classification performance increases. Results of numerical experiments for mach
ine learning problems from UCI repository are presented. Fuzzy classifier design methods comparison with alternative classification methods by performance value demonstrates advantages of the proposed algorithms.
(More)