Safe Reinforcement Learning in Continuous State Spaces

Takumi Umemoto, Tohgoroh Matsui, Atsuko Mutoh, Koichi Moriyama, Nobuhiro Inuzuka

Abstract: In using reinforcement learning [1], [2], there is a problem that states are often provided as continuous rather than discrete values. In this study, we applied continuous state space to reinforcement learning based on the success probability and return (EQ) [3] of conventional research. In addition, we demonstrated that EQ learns more safely than a conventional method in reinforcement learning problem with a continuous state space.