Academy of Mathematics and Systems Science, CAS
Colloquia & Seminars

Speaker:	王曙明，中国科学院大学经济与管理学院
Inviter:
Title:	Data-Driven Production-Service Planning: Ambiguity Aversion, Service-Dependency and Cross-Product Heteroskedasticity
Time & Venue:	2022.11.22 10:30-12:00 腾讯会议：440871514
Abstract:	We consider a joint production-service planning problem with demand uncertainty, where the manufacturer aims to determine the portfolio of products with the associated service levels and the capacity to maximize the expected total profit. We study the problem in a data-driven setting and assume the historical information on demand and the associated covariates are available where the service decision serves as one key covariate for demand. Three salient features of the demand in our problem context are distributional uncertainty (ambiguity), correlation across different products which induces heteroskedasticity for estimation, and service-dependency effect. We first develop a demand prediction model leveraging seemingly unrelated regression (SUR) estimated with feasible generalized least squares (FGLS). The SUR-FGLS model well describes the demand correlation (via capturing the cross-product heteroskedasticity) and estimates the service-dependency effect (via treating service level as a regressor). We then construct an ambiguity set that centers at the SUR-FGLS prediction model to model the demand ambiguity, which leads to a decision-dependent distributionally robust optimization (DRO) model. Statistically, we show our approach enjoys finite-sample performance guarantee and asymptotic consistency, under reasonable (statistical) regularity conditions. Operationally, we identify that the developed data-driven DRO model can be reformulated as a prediction-based empirical optimization model regularized by perceived-shortfall aversion. Exploiting this structure, we analyze the ambiguity-averse operational pattern for the product-service selection. We also analyze the risk-exposure given a product-service decision that captures both distributional- and operational-change effects. Computationally, the proposed model can be reformulated as a mixed-integer conic linear program that enjoys an appealing structure for optimization. Finally, sufficient numerical experiments demonstrate the effectiveness of our framework.