Improving the Discovery Component of Classifier Systems by the application of Estimation of Distribution Algorithms
Joaquín Rivera, Roberto Santana · 2009
Since Holland and Reitman [2] published their CS-1 a great amount of studies concerning classifiers systems (CSs) have appeared. The needs of studying crucial matters in the functioning of CS determined that researchers paid little attention to the evolutionary component of these systems. Nevertheless in the last years GA's have experienced a considerable development that makes necessary to revise an update current applications of EA's to CSs. The goal of our current research is to incorporate to the CS's machinery new Evolutionary Techniques that make of them more powerful machine learning tools. In this paper we introduce an Estimation Distribution Algorithm to perform as the discovery component of the classifier system XCS. Estimation of Distribution Algorithms [1][6][5.5] are based on probability theory and statistics. They are population based optimization methods that use selection. Instead of applying genetic operators to the selected population these algorithm estimate probability distribution of individuals and use this information to generate new points. EDA have shown to be superior [5.5] to classical GA's in the optimization of a wide set of functions. The XCS classifier system[8] has been successful in solving different learning problems. It has deserved a great attention in recent years and settled theoretical basis for the study of generalizations. For the interest reader it may be important to revise [9], [4] and [5] for recent developments concerning XCS. The rules in XCS store three main parameters: prediction pj, prediction error εj, and fitness Fj. The prediction