Reward-Guided Individualised Communication for Deep Reinforcement Learning in Multi-Agent Systems

Broadcasting communication poses a fundamental challenge in Multi-Agent Deep Reinforcement Learning, prompting the emergence of decentralised communication as a promising paradigm. However, the current approach in decentralised communication exhibits drawbacks, including message overhead and asynchronous network updating. More critically, the method predominantly relies on divergence as a metric for communication network updating, which fails to align with the goal of performance maximisation in Reinforcement Learning (RL). To address these limitations, this paper introduces Reward-Guided Individualised Communication (RGIC), a method that integrates rewards into the communication network. By adhering to RL principles, RGIC facilitates purposeful one-to-one interactions and enhances overall performance. The optimised learning process of RGIC leads to accelerated convergence, enhanced efficiency, and reduced computational requirements. Extensive experimentation validates the efficacy of RGIC, establishing its suitability for real-world multi-agent scenarios that demand real-time decision-making and reward-driven actions.