Effective and reliable online classification combining XCS with EDA mechanisms

Abstract

Learning Classifier Systems (LCSs), such as XCS and other accuracy-based classifier systems, evolve a distributed problem solution online. During the learning process, rule quality is assessed iteratively using techniques based on gradient-descent, while the rule structure is evolved using selection and variation operators of evolutionary algorithms. While using standard variation operators suffices for solving some problems, it does not assure an effective evolutionary search in many difficult problems that contain strong interactions between features. Specifically, it was shown that standard crossover operators can frequently disrupt important combinations of features, which often results in poor performance. This chapter describes how advanced EDAs can be integrated into XCS in order to ensure effective exploration even for problems in which features strongly interact and standard variation operators lead to poor XCS performance. In particular, the chapter incorporates the model building and sampling techniques from BOA and ECGA into XCS. The chapter shows that the two proposed algorithms ensure that the solution is found efficiently and reliably. The results presented in this chapter thus suggest that the research on combining standard LCSs with advanced EDAs holds a big promise and represents an important area for future research on LCSs and EDAs. © Springer-Verlag Berlin Heidelberg 2006.