| 主讲人 |
王曙明 |
简介 |
<p>We investigate a two-stage supply-network design problem with random demands from multiple customers, where the decision maker needs to decide which production facilities to open and select emergency supplier (in case of demand shortage) for each customer (first stage) to collectively fulfill the demand (second stage), so as to minimize the expected total cost. We consider the problem in a data-driven setting, where historical information on demand and some covariates are accessible. In practice, determinants for the multi-dimensional random demand over the supply network are usually various, complicated, and even unobservable, which can hardly be well characterized by the observable historical demand and covariates available. To address the demand ambiguity and the underlying unobservable demand effects in a statistically coherent fashion, we develop a predictive distributionally robust approach that incorporates into the ambiguity set a general factor model for capturing both explanatory effects of observable and latent covariates for the demand. We identify an implied asymptotic decision-dependent worst-case distribution, which facilitates our operational analysis on the production and emergency supplies, and leads to a reformulation of mixed-integer conic program for the problem. We also derive finite-sample and asymptotic performance guarantees of the model solution under several statistical regularity conditions. Finally, sufficient experiments using real-life regional demand data with macroeconomic covariates demonstrate the effectiveness of our framework.</p> |
