具有理论保证的多智能体策略互惠（英文版）

英文标题：Multi-agent Policy Reciprocity with Theoretical Guarantee中文摘要：本文提出了一种新的多智能体策略互惠（PR）框架，其中每个智能体可以在不匹配的状态下充分利用跨智能体策略，并定义了一个不匹配状态的邻接空间并设计一个即插即用模块的值迭代，以提高 PR 的可扩展性和稳定性，实验证明 PR 在离散和连续环境中优于现有的各种 RL 和转移 RL 方法。英文摘要：Modern multi-agent reinforcement learning (RL) algorithms hold greatpotential for solving a variety of real-world problems. However, they do notfully exploit cross-agent knowledge to reduce sample complexity and improveperformance. Although transfer RL supports knowledge sharing, it ishyp