Multi-Agent Reinforcement Learning Based on Cross Task Information Sharing
Yu Gao, Lizhong Zhu, Yunting Liu, Jiaming Yang · 2024
This article selects StarCraft as the experimental environment, which has a complex and vast action and state space, making it difficult for intelligent agents to make optimal decisions. In this article, the sharing of parameter experience values among multiple agents in a single task is used to enhance the collaboration ability between agents and avoid repetitive learning of basic strategies. In addition, after distilling and refining the core strategies of individual tasks, policy interaction is carried out in multiple tasks of the same type. By establishing a joint objective function for multiple tasks and alternating optimization for mutual constraints and updating iterations, information exchange between cross task multi-agent systems is completed, solving the problem of difficulty in obtaining effective data in complex scenarios and improving the learning efficiency of micro unit joint decision-making. This paper also adds qualification tracking to solve the problem of reward delay at the initial stage of model training, and uses course transfer learning to complete incremental joint training.