Reward Function by Using Inverse Reinforcement Learning and Fuzzy Reasoning

Abstract

A reward function estimated by inverse reinforcement learning has been used to acquire a robot’s controller. Inverse reinforcement learning requires observed sequences of actions to estimate a reward function. There are few models of the sequence that gives an optimal motion of the robot, therefore a suboptimal one may be given. The suboptimal sequences, however, include some errors and ambiguities. In this paper, we propose a method for quantifying the ambiguity of the reward function, which is designed by inverse reinforcement learning, using fuzzy reasoning.