Inductive operators and rule repair in a hybrid genetic learning system: Some initial results

Abstract

Symbolic knowledge representation schemes have been suggested as one way to improve the performance of classifier systems in the context of complex, real-world problems. The main reason for this is that unlike the traditional binary string representation, high-level languages facilitate the exploitation of problem specific knowledge. However, the two principal genetic operators, crossover and mutation, are, in their basic form, ineffective with regard to discovering useful rules in such representations. Moreover, the operators do not take into account any environmental cues which may benefit the rule discovery process. A further source of inefficiency in classifier systems concerns their capacity for forgetting valuable experience by deleting previously useful rules.

In this paper, solutions to both of these problems are addressed. First, in respect of the unsuitability of crossover and mutation, a new set of operators, specifically tailored for a high level language, are proposed. Moreover, to alleviate the problem of forgetfulness, an approach based on the way some enzyme systems facilitate the repair of genes in biological systems, is investigated.