An Intrinsic Reward Mechanism for Multi-Agent Reinforcement Learning Based on Cooperative Awareness and Attention

The increasing deployment of multi-agent systems in fields such as intelligent transportation and robotic swarms has made enhancing inter-agent cooperation a critical challenge. In this study, a novel framework is proposed to foster cooperative motivation in multi-agent reinforcement learning by introducing counterfactual reasoning to design intrinsic rewards. Central to the framework are two complementary metrics: cooperative awareness, which quantifies an agent’s integration of peer information, and cooperative attention, which measures the bidirectional strategic influence of states on agents. The experimental results indicate that the proposed method achieves notable improvements in learning performance while facilitating effective cooperation among agents. Furthermore, a comprehensive evaluation conducted on the Multi-UAV game-theoretic confrontation platform demonstrates the robustness and efficacy of the approach, highlighting its potential applicability to real-world multi-agent systems.