Reproduce “Quantifying Potential Observation Missingness in Inverse 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
Leo Benac, Abhishek Sharma, Alihan Huyuk, Finale Doshi-Velez
The authors introduce MP-IRL, a method for measuring how much episode-specific context would help a shared reward explain demonstrated actions when recorded observations may omit information. It learns a base reward, then adds localized trajectory-level reward perturbations under a near-optimal-expert assumption. In navigation, decision-region accuracy rises from 67.9% to 99.6%; this measures potential missingness, not recovered variables or proven causes.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.