Proposal and evaluation of the penalty avoiding rational policy making algorithm with penalty level

Abstract

Reinforcement learning (RL) is a kind of machine learning. It aims to adapt an agent to a given environment by utilizing a reward and a penalty. We know the Penalty Avoiding Rational Policy Making algorithm (PARP) [5] and the Penalty Avoiding Profit Sharing (PAPS) [6] as examples of RL systems that are able to suppress a penalty and learn a rational policy. However they cannot treat multiple penalties. In this paper, we extend PARP/PAPS to the environments where there are some kinds of penalties. We propose the Penalty Avoiding Rational Policy Making Algorithm with Penalty Level (PARP L ) that can control how to avoid penalties. We show the effectiveness of PARP L by soccer game simulations.