Genetic Algorithms for Pattern Recognition 论文

摘要

Genetic algorithms (GAs) have emerged as effective search and optimization methods with applications in several problem domains. When the underlying search space has several locally optimal solutions apart from the globally optimal solution (i.e., the search space is multimodal), GAs emerge as worthy alternatives to traditional optimization techniques. For several years since their inception in 1975, GAs have been molded in the form proposed by Holland, characterized by constant control parameters and fixed length encodings. Recent research has led to variations in the basic GA mechanism. New selection, mutation, and crossover strategies, distributed and parallel implementations, and adaptive mechanisms to modify the control parameters have been proposed. The control parameters of a GA-crossover probability, mutation probability, and population size-critically control the performance of GAs. In the last few years, several researchers have experimented with adaptive mechanisms to dynamically vary the control parameters to improve the performance of GAs. The success that they have achieved in their pursuits makes it worthwhile to survey strategies for adapting the control parameters. In this chapter, we briefly review recent work on adaptive strategies for modifying control parameters of GAs. Next we discuss in detail our own efforts in this direction which have led to the genesis of the Adaptive Genetic Algorithm, a very effective GA variant for multimodal optimization.