Information compression effect based on PCA for reinforcement learning agents' communication
Akira Notsu, Katsuhiro Honda, Hidetomo Ichihashi, Ayaka Ido, Yuki Komori · 2012
In general, the amount of required memory and reduction of learning time with explosion of the number of states become problems in reinforcement learning. In this study, as a method of cutting information of the learning table, principal component analysis was used as an information compression method, which is well known. The principal component analysis was applied to the learning table directly. Also, the influence given by reducing the principal component vector extremely when restructuring state space and action space was reviewed. In a numerical experiment, it was confirmed that the proposed method cut the amount of information and without a big change to learning speed and agents' minimal communication can had positive effect on the learning.