Adapting Bagging and Boosting to Learning Classifier Systems

Abstract

Learning Classifier Systems (LCSs) have demonstrated their classification capability by employing a population of polymorphic rules in addressing numerous benchmark problems. However, although the produced solution is often accurate, the alternative ways to represent the data in a single population obscure the underlying patterns of a problem. Moreover, once a population is dominated by over-general rules, the system will sink into the local optimal trap. To grant a problem’s patterns more transparency, the redundant rules and optimal rules need to be distinguished. Therefore, the bagging method is introduced to LCSs with the aim to reduce the variance associated with redundant rules. A novel rule reduction method is proposed to reduce the rules’ polymorphism in a problem. This is tested with complex binary problems with typical epistatic, over-lapping niches, niche-imbalance, and specific-addiction properties at various scales. The results show the successful highlighting of the patterns for all the tested problems, which have been addressed successfully. Moreover, by combining the boosting method with LCSs, the hybrid system could adjust previously defective solutions such that they now represent the correct classification of data.