Reproduce “Off-Policy Evaluation in Partially Observed Markov Decision Processes under Sequential Ignorability” 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
Yuchen Hu, Stefan Wager
Researchers at Stanford University developed a Partial History Importance Weighting (PHIW) estimator for off-policy evaluation in Partially Observed Markov Decision Processes (POMDPs) under sequential ignorability. The method achieves polynomial error rates, outperforming model-free approaches by leveraging hidden Markovian structure, and its rate-minimality is theoretically established.
Key insights
- POMDPs provide a robust framework for OPE, offering more realism than MDPs and more statistical efficiency than model-free approaches, bridging a critical modeling gap.
- The statistical efficiency of OPE in POMDPs is intrinsically linked to the product of the average mixing time and the overlap constant, indicating what system properties dictate problem difficulty.
- Truncating the importance weighting history (PHIW) allows for achieving polynomial error rates, effectively mitigating the curse of dimensionality while retaining approximate unbiasedness due to system mixing properties.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.