Seminar: Data-Driven Objective Selection in Multi-Objective Optimization

Dr. Taewoo Lee

Assistant Professor of Industrial Engineering
University of Houston
Houston, TX
tlee6@uh.edu

Abstract

One of the challenges in multi-objective optimization is to identify which objectives are critical in the optimization in a specific application domain. Currently practitioners typically rely on their own experience and trial-and-error processes to come up with final objectives to use, which is often time-consuming and may lead to suboptimal solutions. We propose a novel approach that infers from data the most critical objectives for a multi-objective optimization problem. Specifically, we propose an inverse-optimization-based approach with a cardinality constraint. We elucidate a relationship between the inverse model and traditional feature selection approaches and analyze the submodularity of the model as a set function. We provide a bound for the greedy algorithm, which generalizes the result by Nemhauser (1978). We compare the greedy algorithm to the exact cardinality-constrained model and show that the greedy algorithm efficiently finds a near-optimal set of objectives. We apply this methodology to radiation therapy treatment planning where one of the biggest bottlenecks is to find clinical ob jectives that can lead to successful treatments.

Biographical Sketch

Dr. Taewoo Lee is an Assistant Professor of Industrial Engineering at the University of Houston. His research focuses on data-driven learning and decision making in healthcare applications including cancer therapy, diabetes treatment, and organ transplantation. Dr. Lee's work has appeared in journals such as Operations Research and Medical Physics, and won several awards including CORS best paper competition and INFORMS Healthcare Application Society best paper competition. He holds a PhD from University of Toronto (2015), and was previously a postdoctoral fellow at Rice University (Computational and Applied Math).

*Light refreshments will be served at 12pm