PEPACS: integrating probability-enhanced predictions to ACS2.

Many real-world environments are non-deterministic, thus presenting learning challenges for Anticipatory Learning Classifier Systems (ALCS). Maze problems have been widely used in ALCS literature, as they provide such environments. However, few ALCS can efficiently run in such mazes, having trouble to build a complete and accurate internal representation of their environment. ALCS can implement Probability-Enhanced Predictions (PEP) in their classifiers. Those PEP permit ALCS to handle non-deterministic environments, but they have never been experimented within ACS2 (Anticipatory Classifier System 2), yet one of the most advanced ALCS, and tested within a thorough maze benchmark. This paper introduces PEPACS, an ALCS that integrates PEP to ACS2. PEPACS integrates an aliasing detection algorithm that let the system build PEP-enhanced classifiers, and the learning process was adjusted so as to avoid over-generalization issues. The presented results show that PEP with ACS2 is a suitable approach to handle at least one kind of non-determinism (namely, the perceptual aliasing issue), without impairing ACS2 functionalities, and that PEP allows the system to build a complete and accurate internal representation of its environment.