Reproduce “Robust Parameter Learning for Uncertain MDPs” 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

Yannik Schnitzer, Alessandro Abate, David Parker

A framework is proposed for learning Uncertain Markov Decision Processes (UMDPs) by explicitly incorporating known parametric dependencies found in Parametric MDPs. This method constructs tighter uncertainty models, leading to more robust policies with stronger certified performance guarantees, demonstrating reduced relative gaps in performance by orders of magnitude on benchmarks compared to independent interval learning.

Key insights

  • Leveraging known parametric dependencies from pMDPs significantly improves the tightness of learned UMDPs compared to traditional methods that assume independent transition uncertainties.
  • Approximations of the intractable coupled uncertainty region, particularly the expression-wise projection, strike an effective balance between model accuracy and computational feasibility for robust policy computation.
  • The framework ensures Probably Approximately Correct (PAC) guarantees, meaning the learned uncertainty models reliably contain the true system dynamics with high confidence.

Summary by alphaXiv.

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