Self-generating method of behavioral evaluation for reinforcement learning among multiple coordinated robots

摘要

In this paper, we present a novel self-generating algorithm for behavioral evaluation, which is used to evaluate self-selected behaviour in a reinforcement learning system. This behavioral evaluation is composed of rewards and self-evaluated standards. Rewards are given by the operator as one of the methods for understanding the purpose of tasks; and self-evaluated standards are obtained as the result of executions. Each robot can generate the evaluation depending on its situations by using the proposed method, and therefore the robots can create cooperative behaviours even if the number of robots or tasks is changed dynamically. We performed simulation experiments to study the effectiveness of the proposed method. The experimental results confirm that each robot can generate evaluations for creating cooperative behaviours without changing the algorithm during the simulation experiments.