Reproduce “Behavior-Discriminative Reward Shaping for Reward-Robust Reinforcement Learning” 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

Zixuan Liu, Fangzheng Wu, Brian Summa, Zizhan Zheng

The authors introduce reward-robust reinforcement learning that removes reward variations equivalent under potential-based reward shaping (PBRS). A state potential defines a shaping term whose discounted contributions preserve policy rankings; the method projects each equivalence class to a canonical reward while preserving the robust objective. On Walker2d, reported return was −188.14 ± 9.26 for the canonical reward and −259.59 ± 16.07 when its removed component was restored, supporting the mechanism.

Summary by alphaXiv.

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