(517b) Data-Driven Discovery of Decision-Making Processes Via Inverse Optimization

Decision making is fundamental to everyday life, and a good understanding of decision-making mechanisms is crucial for predicting the behavior of autonomous agents, learning from experts, and optimizing systems involving various decision makers. But many decision-making processes are unknown or poorly understood. For example, experts in the operation of chemical plants make decisions based on years of experience, but their decision strategies often are not well documented and, due to the complexity of the manufacturing processes, difficult to explain even to fellow operators. As a result, the complete transfer of expert knowledge to new operators remains an unsolved problem.

While we do not know how exactly some decisions are made, we often have access to data in the form of decisions made in the past. Inverse optimization (IO) is an emerging method that can leverage such data to learn unknown decision-making processes. The fundamental idea of IO is to use mathematical optimization as the model for decision-making, where decisions can be viewed as the optimal or near-optimal solutions to an underlying optimization problem. Given observed decisions made by an agent, the goal of IO is to learn the unknown optimization model that best represents the agent’s decision-making process. A key advantage of the IO approach is that it can directly consider constraints, which allows us to leverage all the modeling flexibility of mathematical programming, incorporate domain knowledge, and hence obtain inherently interpretable decision-making models; this makes it distinct from common black-box machine learning methods such as deep learning. In this talk, we review recent advances in data-driven IO and present case studies that demonstrate its efficacy in learning interpretable decision-making models.