抄録
Reinforcement Learning (RL) is one of promissing approaches for controlling an autonomous robot. However, its performance is quite sensitive to the segmentation of state and action spaces. This paper describes an RL mobile robot, the task of which is passing through a narrow short route instead of a wide but long route. The robot needs appropriately segmented state and action spaces to avoid punishments, otherwise the robot tends to fail to pass through the narrow route. In order to overcome this unwanted situation, we apply our proposed technique, named Bayesian-discrimination-function-based Reinforcement Learning (BRL). We investigate the performance of BRL through computer simulations and analyze the learning process.
