Dealing with uncertainty : a comparison of robust optimization and partially observable Markov decision processes

Abstract

Uncertainty is often present in real-life problems. Deciding how to deal with this uncertainty can be difficult. The proper formulation of a problem can be the larger part of the work required to solve it. This thesis is intended to be used by a decision maker to determine how best to formulate a problem. Robust optimization and partially observable Markov decision processes (POMDPs) are two methods of dealing with uncertainty in real life problems. Robust optimization is used primarily in operations research, while engineers will be more familiar with POMDPs. For a decision maker who is unfamiliar with one or both of these methods, this thesis will provide insight into a different way of problem solving in the presence of uncertainty. The formulation of each method is explained in detail, and the theory of common solution methods is presented. In addition, several examples are given for each method. While a decision maker may try to solve an entire problem using one method, sometimes there are natural partitions to a problem that encourage using multiple solution methods. In this thesis, one such problem is presented, a military planing problem consisting of two parts. The first part is best solved with POMDPs and the second with robust optimization. The reasoning behind this partition is explained and the formulation of each part is presented. Finally, a discussion of the problem types suitable for each method, including multiple applications, is provided.