Reproduce “Unifying Goal-Conditioned RL and Unsupervised Skill Learning via Control-Maximization” 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

Alireza Modirshanechi, Benjamin Eysenbach, Peter Dayan, Eric Schulz

The authors develop a theoretical framework linking goal-conditioned reinforcement learning and mutual-information skill learning through goal sensitivity, the effect of commanded goals on future behavior. For persistent-goal and exact-timing tasks, maximizing sensitivity maximizes average performance. For opportunity-window tasks, sensitivity provides a lower bound and regret guarantee, not equivalence. The results assume finite spaces and specified goal and skill distributions.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.