2018 Volume 30 Issue 4 Pages 658-665
A classifier sysem ACSM (ACS 2 with Memory) (Hayashida et al., 2014) is developed on the basis of ACS (Anticipatory Classifier System) (Stolzmann, 1997, 1998) and ACS2 (Butz and Stolzmann, 2002) which is an improved system of ACS. ACSM includes internal memory to distinguish the aliased states of POMDPs (Partially Observable Markov Decision Process) and avoid some wasteful searching process. This paper optimize the complicated procedure of the learning process of ACSM, and aims to reduce unnecessary computation time or improve learning performance. Numerical experiments are executed using some maze problems which are adopted in a number of related papers as benchmark problems of POMDPs including aliased states. The experimental results indicate the proposed method can find the solutions more efficiently compared to ACS2 and ACSM.
