Locally Centralized Execution for Less Redundant Computation in Multi-Agent Cooperation

Abstract

Decentralized execution is a widely used framework in multi-agent reinforcement learning. However, it has a well-known but neglected shortcoming, redundant computation, that is, the same/similar computation is performed redundantly in different agents owing to their overlapping observations. This study proposes a novel method, the locally centralized team transformer (LCTT), to address this problem. This method first proposes a locally centralized execution framework that autonomously determines some agents as leaders that generate instructions and other agents as workers to act according to the received instructions without running their policy networks. For the LCTT, we subsequently propose the team-transformer (T-Trans) structure, which enables leaders to generate targeted instructions for each worker, and the leadership shift, which enables agents to determine those that should instruct or be instructed by others. The experimental results demonstrated that the proposed method significantly reduces redundant computations without decreasing rewards and achieves faster learning convergence.