Seminar Series | Graduate Student Colloquium
270 Journalism Building
242 W 18th Ave
Columbus, OH 43210
United States
Presenter: Chennan Zhou
Faculty advisor: Guzin Bayraksan
Committee members: Guzin Bayraksan, Cathy Xia, Sam Davanloo
Title of my research: Effective Scenarios in Distributionally Robust Optimization: Properties and Acceleration of Decomposition Algorithms
Brief description: Classical stochastic optimization assumes that the decision maker has complete knowledge of the underlying distribution and is risk neutral. In contrast, classical robust optimization assumes that the decision maker has no knowledge of the uncertainty except for its possible range and can be very conservative. Distributionally robust optimization (DRO) is an alternative to classical stochastic and robust optimization approaches. DRO assumes that the underlying distribution of uncertainty is unknown but lies in an ambiguity set of distributions—a family of distributions consistent with the prior knowledge about uncertainty. DRO finds the worst-case expected cost among all distributions within the ambiguity set. This approach is more realistic because many real-world problems have some data but not all uncertainties are fully known.
This dissertation focuses on effective scenarios for DRO, which are the scenarios that alter the optimal value if removed from the problem. This dissertation explores effective scenarios with a general ambiguity set, and specifically delves into the ambiguity sets formed by the Cressie-Read family of power divergences (DRO-CR) and the Wasserstein distance (DRO-W). This dissertation also studies the computational benefits of effective scenarios in large-scale problems by accelerating decomposition algorithms to solve this class of problems.