Abstract: Multi-agent decision-making offers extensive engineering applications,particularly in the cooperative control tasks.Po-licy gradient-based reinforcement learning methods,which directly model policy distributions,are more conducive to exploring diverse strategies in complex reward scenarios.These methods also demonstrate consistently high empirical efficiency across both discrete and continuous action spaces.Although parameter-sharing mechanisms are widely adopted in policy gradient frameworks to improve convergence efficiency for collaborative tasks,the lack of attention to action semantic modeling introduces critical limitations,especially in mitigating action homogenization among agents.To solve this issue,this paper proposes CSF method from a graph-based modeling perspective.The CSF framework employs a graph autoencoder to learn correlation-aware semantic embeddings within the action space,subsequently achieving information fusion through dynamic integration of agent-specific beha-vioral features with semantic embeddings.This fusion mechanism aggregates collaborative behavioral information into agent-specific latent representations,enabling interdependent policy space exploration across agents.Comprehensive experiments conducted on diverse complex task scenarios within the StarCraft and Google Research Football environments demonstrate that CSF achieves superior performance over state-of-the-art algorithms,thus validating its effectiveness in facilitating inter-agent collaboration.

Key words: Multi-agent reinforcement learning, Graph autoencoder, Semantic relations, Feature fusion, Behavior decision-making