UC Berkeley

The past decade has seen tremendous growth in the availability of voluminous high-quality data in many modern decision systems such as supply chain management. Big data provides new opportunities to tackle one of the main difficulties in decision-making systems – uncertain behavior driven by some unknown ground truth probability distribution. My dissertation mainly focuses on answering the following question: How do we use (non-perfect) data samples generated from the unknown distribution to make more automatic and reliable decisions, especially for supply chain management and retail operations?

In Chapter 2, this dissertation study a practical multi-period inventory management problem faced by online retailers and propose a practical, data-driven, end-to-end (E2E) framework for multi-period inventory management that leverages the power of deep learning. Instead of adopting the two-step approach that first forecasts demand and vendor lead-time (VLT) separately and then, based on the predictions, solves for a solution that minimizes inventory costs, we directly learn the underlying mapping from features to the optimal solution by adopting a supervised learning approach.

In Chapter 3, we propose an integrated conditional estimation-optimization (ICEO) framework that uses statistical learning theory to estimate the underlying conditional distribution while considering the structure of the optimization problem. This method provides a generic learning framework and applies to a broad class of convex contextual stochastic optimization problems with uncertainty in the objective. We provide statistical performance guarantees and address the computational challenges that arise in performing this framework.

In Chapter 4, we focus on data-driven distributionally robust optimization (DRO) methods. We examine the conditional quantile prediction problem, which is equivalent to the contextual newsvendor problem, and propose a DRO framework. In contrast to the traditional setting where the design points are considered as random and i.i.d., we consider a fixed-design setting, which is more suitable for cases when features are pre-designed, such as through a series of controlled experiments.

Chapter 5 continues our study on the contextual newsvendor problem and considers inter-temporal correlations and moderate non-stationarities in the contextual demand process. Under these comparatively more realistic assumptions, we provide performance guarantees in the form of out-of-sample generalization bounds.