Policy-Based Branch-And-Bound For Infinite-Horizon Multi-Model Markov Decision Processes

Markov decision processes (MDPs) are models for sequential decision-making that inform decision making in many fields, including healthcare, manufacturing, and others. However, the optimal policy for an MDP may be sensitive to the reward and transition parameters which are often uncertain because parameters are typically estimated from data or rely on expert opinion. To address parameter uncertainty in MDPs, it has been proposed that multiple models of the parameters be incorporated into the solution process, but solving these problems can be computationally challenging. In this article, we propose a policy based branch-and-bound approach that leverages the structure of these problems and numerically compare several important algorithmic designs. We demonstrate that our approach outperforms existing methods on test cases from the literature including randomly generated MDPs, a machine maintenance MDP, and an MDP for medical decision making. (C) 2020 Elsevier Ltd. All rights reserved.