Cooperative Game-Theoretic Credit Assignment for Multi-Agent Policy Gradients via the Core

Researchers propose CORA, a new cooperative game-theoretic method for credit assignment in multi-agent reinforcement learning that uses coalition-wise advantage allocation. The approach addresses policy optimization challenges by evaluating marginal contributions of different agent coalitions and demonstrates superior performance across various benchmarks.

• →CORA uses cooperative game theory's core allocation to distribute advantages among agent coalitions rather than individual agents.
• →The method combines clipped double Q-learning to mitigate overestimation bias in coalition-wise advantage estimation.
• →Random coalition sampling is employed to reduce computational overhead while approximating core allocation efficiently.
• →Experiments show CORA outperforms baseline methods across matrix games, differential games, and multi-agent collaboration benchmarks.
• →The research highlights the importance of coalition-level credit assignment for advancing coordinated behavior in multi-agent systems.