Mamba-Driven Strategy for Multi-Agent Games: Towards a Stable Cooperative Value Framework

Multi-agent reinforcement learning (MARL) has shown great potential in addressing collaborative issues within multiagent systems (MAS), particularly in domains such as autonomous driving, robotic collaboration, and network systems. This study introduces an innovative MARL algorithm, MambaCoV, which focuses on tackling the challenges of complex interactions between agents and adapting to dynamic environments. MambaCoV employs a constrained communication framework that facilitates the exchange of essential information among agents, allowing the coordination of actions and the collective optimization of overall performance. Through this mechanism, agents can fairly assess each other's contributions and allocate rewards accordingly, thereby reducing the variance in policy gradient estimation and enhancing the learning efficiency of multi-agent systems. In policy learning, we introduce a network architecture centered around the Mamba model, designed to effectively utilize historical information. Experimental evaluations in cooperative navigation and Starcraft Multi-Agent Challenge (SMAC) demonstrate that MambaCoV exhibits faster adaptability and more stable performance improvements in the early stages of training compared to existing advanced MARL methods. This result confirms that effective communication and coordination between agents are key factors in improving performance. Moreover, through precise credit assignment, MambaCoV ensures that each agent's contributions are reasonably assessed, enhancing the efficiency and stability of team collaboration.