通信受限下的协作式多智能体强化学习方法

打开文本图片集

中图分类号：TP18 文献标志码：A 文章编号： 1000-5013（2026）02-0193-09

Abstract：To address the problem that most existing collaborative multi-agent reinforcement learning methods adopt overly idealized communication assumptions，a cooperative multi-agent reinforcement learning methods under communication constraints is proposed. First，a more realistic communication constrained environment is constructed by introducing random information loss and additive Gaussian white noise disturbance. Then，a residual connection-based value decomposition method is proposed，leveraging residual structures to enhance the robustness of system against communication quality fluctuations and observational noise. Finaly， the proposed method is validated in a communication constrained test environment built on the StarCraft multi -agent challenge benchmark. Experimental results show that the proposed method performs excellently under various communication-constrained scenarios，significantly outperforming current mainstream multi-agent reinforcement learning methods.

Keywords：communication constraint；cooperative multi-agent reinforcement learning；residual connection; value decomposition

随着深度学习和多智能体系统的发展，多智能体强化学习（MARL）近些年受到了广泛关注[1-3]，并已被应用于解决各种协作式多智能体决策任务，例如，自动驾驶[4]、编队协同控制[5]及城市交通网络拥堵协同控制[6-7]等。（剩余11773字）