Meta-Strategy Learning by using Agent Simulation of Collision Avoidance

Abstract

In our cooperative behavior, there are two type strategies passive behavioral strategies based on others' behavior and active behavioral strategies that oneself acts first. In order to realize a robot that can communicate with human, it is necessary to acquire such behavioral strategies. However, it is not clear how to acquire meta-strategy to switch those strategies. In this study, we have experimented with multi-agent collision avoidance simulations as an example of cooperative tasks. In the experiment, we have used reinforcement learning to obtain an active and passive strategy by rewarding the interaction with agents facing each other. Additionally, we have acquired meta-strategy by reinforcement learning to selectively use those strategies and evaluated them.