Why state differentiation in ACS2 is not enough in aliased environments
Mateusz Łabędzki, Olgierd Unold · Proceedings of the Genetic and Evolutionary Computation Conference Companion · 2025
This study investigates recent advances in Anticipatory Learning Classifier Systems (ALCSs) designed to address the aliasing problem that arises when multiple environmental states produce identical perceptions. Four ALCS-based algorithms, ACS2, PEPACS, BACS, and BEACS, are described and rigorously evaluated. The performance of these algorithms is assessed using a comprehensive set of 26 aliased and non-aliased maze environments, with an in-depth analysis of the results. Despite the significant progress made by the BEACS algorithm in mitigating the aliasing issue, several limitations persist, including suboptimal decision policies and model aliasing, which occurs when the algorithm produces overly general rules. These shortcomings are critically examined, with illustrative examples and quantitative metrics provided to underscore the need for further refinement.