Reproduce “Modeling Others using Oneself in 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
Roberta Raileanu, Emily Denton, Arthur Szlam, Rob Fergus
Researchers at NYU and Facebook AI Research introduce Self Other-Modeling (SOM), a framework where agents infer the hidden goals of others by leveraging their own policy network. This approach consistently outperforms prior baselines in multi-agent cooperative and adversarial tasks, with agents accurately inferring 60-80% of opponent goals within early game steps.
Key insights
- Human cognitive theories, specifically "theory of mind" and "simulation theory," suggest that understanding others involves simulating one's own mental processes from their perspective.
- An agent's own learned policy provides a robust and parameter-efficient model for inferring the hidden goals of other agents, especially when differences in behavior stem primarily from differing objectives.
- Explicitly inferring hidden goals enables more sophisticated cooperative and competitive strategies compared to merely predicting other agents' actions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.