Reproduce “Mix-ME: Quality-Diversity for Multi-Agent Learning” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Garðar Ingvarsson, Mikayel Samvelyan, Bryan Lim, Manon Flageat, Antoine Cully, Tim Rocktäschel

Mix-ME introduces a Quality-Diversity approach for cooperative multi-agent learning under partial observability, extending MAP-Elites to discover a diverse collection of high-performing multi-agent policies. The method outperforms naive multi-agent baselines in several continuous control tasks and enhances robustness to leg dysfunction, demonstrating that decentralized control coupled with diversity discovery can yield more adaptive multi-agent systems.

Key insights

  • Applying Quality-Diversity principles to multi-agent learning, specifically with the proposed team-crossover operator, effectively discovers diverse and high-performing multi-agent policies.
  • Decentralized control combined with diversity-seeking methods can lead to more robust multi-agent systems, particularly against agent-specific failures like leg dysfunction.
  • The performance advantages observed in multi-agent QD are not merely an artifact of increased total parameter count, indicating an intrinsic benefit of this approach.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.