Reproduce “Hindsight policy gradients” 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
Paulo Rauber, Avinash Ummadisingu, Filipe Mutz, Juergen Schmidhuber
Hindsight Policy Gradients (HPG) adapts policy-gradient learning to sparse-reward tasks by using trajectories collected for one goal to learn about alternative goals. It evaluates rewards for alternative goals and corrects policy-gradient contributions for action likelihood. In simulated FetchPush, HPG’s average performance was 13.15, versus 4.42 for GCPG at batch size 16; tests covered the authors’ environments, not held-out goals.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.