Credit assignment in heterogeneous multi-agent reinforcement learning for fully cooperative tasks

Credit assignment poses a significant challenge in heterogeneous multi-agent reinforcement learning (MARL) when tackling fully cooperative tasks. Existing MARL methods assess the contribution of each agent through value decomposition or agent-wise critic networks. However, value decomposition techniques are not directly applicable to control problems with continuous action spaces. Additionally, agent-wise critic networks struggle to differentiate the distinct contributions from the shared team reward. Moreover, most of these methods assume agent homogeneity, which limits their utility in more diverse scenarios. To address these limitations, we present a novel algorithm that factorizes and reshapes the team reward into agent-wise rewards, enabling the evaluation of the diverse contributions of heterogeneous agents. Specifically, we devise agent-wise local critics that leverage both the team reward and the factorized reward, alongside a global critic for assessing the joint policy. By accounting for the contribution differences resulting from agent heterogeneity, we introduce a power balance constraint that ensures a fairer measurement of each heterogeneous agent’s contribution, ultimately promoting energy efficiency. Finally, we optimize the policies of all agents using deterministic policy gradients. The effectiveness of our proposed algorithm has been validated through simulation experiments conducted in fully cooperative and heterogeneous multi-agent tasks.