Key points are not available for this paper at this time.
ABSTRACT Multi‐agent reinforcement learning (MARL) often suffers from low sample efficiency and limited behavioral diversity, leading to policy homogenization, insufficient exploration, and reduced robustness. To address these challenges, we propose MADECM, a curiosity‐augmented evolutionary framework built upon MADDPG that integrates curiosity‐driven updates with evolutionary quality‐diversity optimization. MADECM employs random network distillation (RND) to estimate the novelty of each agent's local observations and uses the resulting novelty signal to dynamically allocate additional update frequencies, thereby emphasizing exploration‐relevant experience during training. In addition, MADECM combines population‐based diversification with a quality‐diversity (QD) archive through a staged optimization procedure, enabling the joint improvement of task return and policy diversity. We evaluate MADECM on the multi‐agent particle environment (MPE), including Spread and Reference, which capture cooperative and partially observable dynamics, and on google research football (GRF), which emphasizes long‐horizon sequential decision‐making. Results show that MADECM consistently outperforms strong MADDPG‐based baselines. The modular design of MADECM, consisting of RND‐based novelty estimation and staged QD optimization, further supports consistent generalization across these structurally distinct environments without task‐specific hyperparameter tuning.
Wu et al. (Fri,) studied this question.