Reproduce “Descent-Guided Policy Gradient for Scalable Cooperative Multi-Agent 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
Shan Yang, Yang Liu
DG-PG uses analytical system-state references to guide cooperative multi-agent policy updates, combining agent-specific local directions with reward-based gradients. It estimates whether each action moves affected state quantities toward a policy-independent reference. Under stated assumptions, its minimum-variance guidance weight yields per-agent gradient-estimator variance of O(1), rather than O(N). DG-PG reached its convergence threshold in fewer than 10 episodes on average in simulations with up to 1,500 agents.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.