Reproduce “Learning to Resolve Alliance Dilemmas in Many-Player Zero-Sum Games” 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
Edward Hughes, Thomas W. Anthony, Tom Eccles, Joel Z. Leibo, David Balduzzi, Yoram Bachrach
DeepMind researchers demonstrate that independent reinforcement learning agents fail to form beneficial alliances in many-player zero-sum games due to inherent 'alliance dilemmas'. They show that augmenting agents with learnable contract mechanisms allows them to successfully discover and commit to alliances, leading to improved collective outcomes for the allied players.
Key insights
- Alliance dilemmas, where cooperation is collectively beneficial for a subset of players but individually risky, are a common structural feature in many-player zero-sum games.
- Standard multi-agent reinforcement learning (MARL) algorithms, designed for independent agents, consistently fail to resolve these alliance dilemmas, converging to sub-optimal individualistic behaviors.
- Explicit, learnable contract mechanisms, inspired by economic theory, enable self-interested reinforcement learning agents to overcome individual incentives and successfully form dynamic alliances in competitive multi-player settings.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.