Reproduce “A Separation Principle for Cooperative Multi-Agent Reinforcement 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

Lucia Pezzetti, Nicolas Lanzetti, Antonio Terpin, Florian Dörfler, Giorgia Ramponi

The authors introduce SALT, which learns a shared target-conditioned controller and matches agents to destinations using predicted single-agent costs. Its transport cost upper-bounds population cost-to-go under stochastic dynamics and is exact when dynamics are deterministic. In tests, SALT-PPO reduced average travel time from 56.2 to 38.3 versus MAPPO in one grid setting; Manhattan gains compare only with shortest-path routing.

Summary by alphaXiv.

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