🤖 AI Summary
In multi-agent policy gradient (MAPG) methods, relative overgeneralization (RO)—where joint policies converge to suboptimal solutions due to mutual overfitting among agents—remains a critical challenge.
Method: This paper introduces, for the first time, an optimistic update mechanism into the MAPG framework, coupled with a simple yet effective advantage-truncation-based clipping technique. The method mitigates agents’ excessive dependence on others’ suboptimal behaviors, thereby fostering coordinated exploration. We provide rigorous theoretical analysis proving convergence to Nash equilibrium fixed points.
Results: Evaluated on 19 tasks across Multi-agent MuJoCo and Overcooked benchmarks, our approach outperforms state-of-the-art baselines on 13 tasks and matches them on the remaining 6. These results empirically validate both the theoretical guarantees and practical robustness of the proposed method.
📝 Abstract
*Relative overgeneralization* (RO) occurs in cooperative multi-agent learning tasks when agents converge towards a suboptimal joint policy due to overfitting to suboptimal behavior of other agents. No methods have been proposed for addressing RO in multi-agent policy gradient (MAPG) methods although these methods produce state-of-the-art results. To address this gap, we propose a general, yet simple, framework to enable optimistic updates in MAPG methods that alleviate the RO problem. Our approach involves clipping the advantage to eliminate negative values, thereby facilitating optimistic updates in MAPG. The optimism prevents individual agents from quickly converging to a local optimum. Additionally, we provide a formal analysis to show that the proposed method retains optimality at a fixed point. In extensive evaluations on a diverse set of tasks including the *Multi-agent MuJoCo* and *Overcooked* benchmarks, our method outperforms strong baselines on 13 out of 19 tested tasks and matches the performance on the rest.
