Reproduce “Unified PAC-Bayesian Study of Pessimism for Offline Policy Learning with Regularized Importance Sampling” 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

Imad Aouali, Victor-Emmanuel Brunel, David Rohde, Anna Korba

The authors develop a shared PAC-Bayesian framework comparing regularized importance weights for offline policy learning. It provides bounds and learning objectives across several weight transformations, accounting for bias, variability, and distance from a prior. In simulated image-classification tasks, clipping performs comparably to exponential smoothing under the heuristic, while harmonic regularization is lower; these results do not establish real-world performance.

Summary by alphaXiv.

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