Reinforcement learning for multi-agent with asynchronous missing information fusion method

Most current research on multi-agent reinforcement learning assumes a reliable environment where agents have globally accurate observations. However, this assumption places high demands on communication between agents and has hindered the large-scale application of reinforcement learning algorithms in realistic scenarios. This paper presents, for the first time, a state estimation method that addresses the problem of training failure caused by asynchronous missing information. Our approach addresses the training dispersion problem caused by information asynchrony, noise, measurement errors, and missing information during multi-agent interaction. We use neural networks to construct recursive state models for the agent environment with limited state information and fill in the missing and asynchronous information with state estimates for reinforcement learning training. In the simulation validation stage, we combine the designed asynchronous missing information fusion algorithm with multi-agent reinforcement learning algorithms for validation. Our results demonstrate that our algorithm overcomes the problem of unsuccessful training due to asynchronous information when the information missing rate does not exceed 30%.