Simplifying communication control: a cooperative multi-agent reinforcement learning framework based on group decision-making

Multi-agent cooperative systems face fundamental challenges in distributed learning due to environmental non-stationarity and partial observability. While existing approaches employ communication mechanisms or hybrid action spaces to enhance collaboration, they often rely on complex policy networks processing state features from limited-bandwidth messaging. We propose a simplified cooperative multi-agent reinforcement learning framework that directly generates coordinated actions through customized communication protocols. Building on independent deep Q-learning, our method introduces three key components: i) a structured unidirectional communication network with normalized message formats, ii) a relationship mapping mechanism converting individual decisions into collective behaviors, and iii) group decision-making rules for consensus formation through opinion aggregation. Experimental evaluations across varying cooperation intensities demonstrate superior convergence performance and higher expected discounted rewards compared to baseline methods, particularly in large-scale multi-agent scenarios. The framework’s clear consensus generation and stable group architecture address critical challenges in distributed coordination while maintaining computational efficiency. These advancements offer conceptual and practical contributions for scalable cooperative systems in intelligent transportation and automated production domains.