Reproduce “Imitation Bootstrapped 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
Tony H. Zhang, Connor Mattson, Daniel S. Brown
IBMARL makes imitation policies participants in multi-agent reinforcement learning. Agents are taught one at a time, then centralized critics arbitrate between imitation and reinforcement-learning actions; a critic-gated penalty encourages actors to retain imitation actions critics prefer. On Buzzwire in simulation, actors surpassed the frozen imitation policy after a median 104,000 interaction steps, versus 296,000 for RLfD and RFT; this is not a real-world result.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.