Reinforcement Learning for Dynamic Environment: Study on Adaptation by Partially Enlarging of the Calibration of State and Action Spaces

Abstract

In general, it is not easy to put Reinforcement Learning into practical use. Our approach mainly deals with the problem of designing state and action spaces. Recently, we have proposed a co-construction method of state and action spaces. In this paper, we propose a detection method of environmental changes and a co-construction method by partially enlarging of the calibration of state and action spaces to adapt dynamic environment. In addition, the validity of the proposed method is confirmed through computational experiments using a so-called "path plannning problem" under dynamic environment.