XCS: Is Covering All You Need?
Connor Schönberner, Sven Tomforde · Proceedings of the Genetic and Evolutionary Computation Conference Companion · 2024
The XCS Classifier System (XCS) can be considered as the most popular and most researched Learning Classifier System (LCS) and was originally intended to learn Reinforcement Learning (RL) problems. However, over time it became clear that the system suffers under issues limiting its performance in multi-step RL problems. This paper provides empirical insights on how the Genetic Algorithm (GA) and the covering mechanism of XCS contribute to the performance in multi-step RL environments, by experimenting with mazes and the Frozen Lake environment. It meets this objective by testing the hypothesis that XCS is able to learn these environments only using covering without activating the GA when using neural prediction or Recursive Linear Least Squares (RLS) prediction. The experiments affirm this hypothesis for neural prediction in several instances, while XCS with linear RLS prediction appears to require the GA. When using neural prediction, the GA of XCS produces considerably more rules than necessary for small performance gains at best or increased instability at worst.