An Empirical Analysis of Action Map in Learning Classifier Systems

An action map is one of the most fundamental options in designing a learning classifier system (LCS), which defines how LCSs cover a state action space in a problem. It still remains unclear which action map can be adequate to solve which type of problem effectively, resulting in a lack of basic design methodology of LCS in terms of the action map. This paper attempts to empirically conclude this issue with an intensive analysis comparing different action maps on LCSs. From the analysis on a benchmark classification problem, we identify a fact that an adequate action map can be determined depending on a type of problem difficulty such as class imbalance, more generally, a complexity of classification or decision boundary of problem. We also conduct an experiment on a human activity recognition task as a real world classification problem, and then confirm that a suggested adequate action map from the analysis enables an LCS to improve on the performance. Those results claim that the action map should be selected adequately in designing LCSs in order to improve their potential performance