HGAP: Boosting Permutation Invariant and Permutation Equivariant in Multi-Agent Reinforcement Learning via Graph Attention Network

In Reinforcement Learning, deep neural networks play a crucial role, especially in Multi-Agent Systems. Owing to information from multiple sources, the challenge lies in handling input permutations efficiently, causing sample inefficiency and delayed convergence. Traditional approaches treat each permutation source as individual nodes for inference. Our novel approach integrates an attention mechanism, allowing us to capture temporal dependencies and contextually align inputs. The attention mechanism enhances the alignment process, allowing for improved information processing. Empirical evaluations on SMAC environments demonstrate superior performance compared to baselines, achieving a higher win rate on 68% of test evaluations.