Can a classifier system achieve near-optimal performance in complex environments? This research explores the incorporation of Wilson's bit-register memory scheme into the XCS classifier system within non-Markov environments. The system is designed to handle situations where the current state does not fully predict future states. Two key extensions were critical for achieving near-optimal performance in difficult environments. The first was an exploration strategy where external actions were probabilistic but internal “actions” were deterministic. The second involved using a register with more bit-positions than necessary to resolve environmental aliasing. The study discusses the origins and effects of these two extensions, demonstrating their impact on the classifier's ability to learn and adapt. These findings contribute to the advancement of classifier systems and their ability to solve complex problems in real-world scenarios. The insights gained could be valuable in designing more robust and adaptable AI agents for applications such as robotics, control systems, and data mining, enhancing performance in environments with incomplete information.
Published in Evolutionary Computation, this paper fits into the journal's focus on adaptive systems and machine learning. The use of a classifier system, a type of evolutionary algorithm, is consistent with the journal's scope. The paper contributes to the understanding of how evolutionary computation methods can be applied to solve problems in complex and uncertain environments, which is of interest to the journal's readership.