New LP Representations and Algorithms for Robust Markov Decision Processes

This paper introduces new linear programming (LP) representations and strongly polynomial algorithms for robust Markov decision processes (RMDPs). The research focuses on RMDPs with uncertainty in rewards and transitions, specifically within rational polyhedral state-action rectangular uncertainty. By encoding robust policy-iteration steps, the study constructs a single LP that can recover optimal robust values and policies. The paper also provides a general complexity analysis for robust policy iteration, leading to improved complexity bounds for various RMDP types and establishing new strongly polynomial bounds for general interval, weighted $\ell_1$, and Wasserstein RMDPs, as well as turn-based stochastic games. AI

IMPACT Introduces novel algorithmic approaches for sequential decision-making under uncertainty, potentially impacting AI systems that require robust planning.

RANK_REASON The item is an academic paper detailing new algorithms and representations for a specific type of decision-making process. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• arXiv
• $\ell_1$ RMDPs
• $\ell_\infty$ RMDPs
• linear programming
• Markov decision processes
• RMDPs
• Wasserstein RMDPs

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →