Proceedings of the Annual Meeting of the Cognitive Science Society

Communication is essential for enhancing the cognition and cooperation of agents in multi-agent reinforcement learning (MARL). However, existing methods often rely on predefined and rigid cognitive patterns, which cannot adapt to dynamic environmental changes and complex inter-agent interactions. In this work, we introduce the Hierarchical Cognitive Graph Autoencoder (HCGA), an adaptive framework that addresses these limitations. HCGA represents inter-agent messages as nodes in a graph with learnable edges, employs a grouping mechanism to integrate related local information into compact latent representations, and then applies hierarchical aggregation to construct a comprehensive global cognition. This approach effectively distills essential information and adaptively uncovers cognitive patterns from dynamic environments, thereby enhancing the overall robustness and efficiency of cognitive processing in MARL tasks. Experimental results demonstrate that HCGA significantly outperforms state-of-the-art methods across various MARL tasks, highlighting its robustness, adaptability, and efficiency.