Reproduce “Towards Optimal Policy Improvement” 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
Yaniv Oren, Viliam Vadocz, Wiktor Zabka, Thomas Evers, Jan Robine, Wendelin Böhmer, Matthijs T. J. Spaan, Martha White and 2 more
The authors introduce I_opt, a policy-improvement operator that scales each action’s update by uncertainty in its estimated advantage while retaining preference for the current policy. It optimizes a tractable certainty-equivalent objective under Gaussian-noise assumptions. In 9×9 Go, it scored 277 Bayes Elo above GumbelAlphaZero with an action-value critic after 34 million training frames, averaged across five seeds.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.