Reproduce “Cooperative Multi-Agent Reinforcement Learning with Submodular Trajectory Rewards” 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

Yasaman Zolfimoselo, Nikolas Geroliminis, Maryam Kamgarpour

The authors introduce GWPO, a method for cooperative planning with trajectory-dependent submodular rewards and decoupled agent dynamics. Each agent uses a sliding window of recent visits, while agents are optimized sequentially using sampled trajectories of those planned. A curvature-based theorem bounds approximation loss. In traffic simulations, memory improves returns over GPO for non-additive rewards, but comparisons with SubPO are mixed.

Summary by alphaXiv.

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