Distributed reinforcement learning that emerges cooperative behavior and communication in heterogeneous multi-agent environments

Abstract

This paper approaches problem-solving in multi-agent environments using deep reinforcement learning. In this paper, we solve a cooperative air rescue task by a fixed-wing aircraft and a helicopter as a problem in multi-agent environments. They have different abilities about speed and expected tasks. Therefore, the purpose of this research is to emerge teamwork that takes advantage of different abilities in multi-agent environments. For this purpose, this paper proposes a method for agents to learn to communicate. The proposed method improves the achievement rate of the cooperative task by transmitting the appropriate communication from a fixed-wing aircraft to a helicopter. We compare the "proposed method," "no communication", and " definite communication " using an air rescue task. From the experiments, we confirm the emergence of the cooperative task by the proposed method and the effectiveness of the proposed method.