Improving Learning Classifier Systems by a Bit-Flip based Local Search

摘要

In this paper we investigate the application of a bit-flip based local search procedure to the evolutionary module of a rule-based Classifier System adopting the Pittsburgh approach, in order to improve both the quality of the classification and the convergence of the evolutionary algorithm. Applying local search to chromosomes has already shown its advantages in improving efficiency of evolutionary processes, especially for what the exploitation task is concerned. A drawback of this approach concerns the adjunctive computational cost of each chromosome evaluation. Our experimental results show how a simple neighbor search speeds up the convergence of the algorithm, resulting in a smaller number of evaluations needed to find a good classifier. Moreover, the quality of the final classifier, measured with a prediction accuracy parameter, is better when the local search is applied.