Self task decomposition for modular learning system through interpretation of instruction by coach

Abstract

One of the most formidable issues of RL application to real robot tasks is how to find a suitable state space, and this has been much more serious since recent robots tends to have more sensors and the environment including other robots becomes more complicated. In order to cope with the issue, this paper presents a method of self task decomposition for modular learning system based on self-interpretation of instructions given by a coach. The proposed method is applied to a simple soccer situation in the context of RoboCup. © Springer-Verlag Berlin Heidelberg 2006.