Learning Diagnostic Rules with Genetic Algorithms | Concepts, Techniques, and Experiences

The paper describes an inductive learning environment called DELVAUX for classiication tasks that learns PROSPECTOR-style, Bayesian classiication rules from sets of examples. A genetic algorithm approach is used for learning Bayesian rule-sets, in which a population consists of sets of rule-sets that generate oospring through the exchange of rules, permitting tter rule-sets to produce oospring with a higher probability. The various genetic operators of our learning environment, such as crossover, mutation, inversion and selection operators, are described. A bucket brigade algorithm for Bayesian rule-sets, called reward punishment mechanism, is introduced that evaluates the performance of a Bayesian rule within a rule-set, and modiied genetic operators that take advantage of such knowledge are introduced. Moreover, heuristics to generate rules more intelligently during the genetic algorithm search process are discussed. Finally, we explore the use of multi-rule-set decision making strategies to improve the learning performance of DELVUAX. We also present empirical results we obtained by running the various versions of the DELVAUX for the classiication of Iris owers, of soybean diseases, of diierent kinds of glass, and for the interpretation of radar images, and the performance of our learning environment is compared with a learning environment that uses neural networks. bucket brigade algorithms for fuzzy rules, comparison between neural networks and genetic algorithms.