Reproduce “Counterfactual Shapley Credit Assignment” 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
Mingxuan Li, Kai-Zhan Lee, Elias Bareinboim
The method assigns reinforcement-learning credit by replaying trajectories with selected actions replaced by baseline-policy choices, then averaging action contributions with Shapley values. Its redistributed rewards theoretically preserve the optimal policy. In the Fork MDP, ϕ-PPO reached 100% success, versus about 58% for REINFORCE and 5% for PPO; this controlled result assumes simulator access and independent exogenous noise.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.