Mining Optimal Policies: A Pattern Recognition Approach to Model Analysis

This paper studies the application of machine learning (ML) algorithms for discovering structural properties of optimal policies in numerically obtained solutions to optimization problems. We propose a framework based on ML for conducting model analysis in a systematic way that complements theoretical and numerical methods. As a proof of concept, we apply the framework to core operations problems, such as inventory management, queuing admission control, multiarmed bandit (MAB), and revenue management problems. We demonstrate how this approach can be used to identify optimal threshold-based policies (inventory management and admission control) and index policies (MAB), as well as for developing new heuristics for revenue management problems. For the MAB problem, our approach leads to a new efficient algorithm for computing optimal index policies. The main contribution of this work is methodological, in proposing and demonstrating the potential of using ML algorithms to analyze optimization problems and devise interpretable policies.