Exact First-Choice Product Line Optimization

• 成果类型: Article
• 署名作者: Bertsimas, Dimitris; Misic, Velibor V.
• 署名单位: Massachusetts Institute of Technology (MIT); Massachusetts Institute of Technology (MIT); University of California System; University of California Los Angeles
• 刊物名称: OPERATIONS RESEARCH
• ISSN/ISSBN: 0030-364X
• DOI: 10.1287/opre.2018.1825
• 发表日期: 2019
• 页码: 651-670
• 关键词: assortment optimization; Benders decomposition; revenue management; selection problem; logit model; CHOICE; DESIGN; algorithm

摘要

A fundamental problem faced by firms is that of product line design: given a set of candidate products that may be offered to a collection of customers, what subset of those products should be offered to maximize the profit that is realized when customers make purchases according to their preferences? In this paper, we consider the product line design problem when customers choose according to a first-choice rule and present a new mixed-integer optimization formulation of the problem. We theoretically analyze the strength of our formulation and show that it is stronger than alternative formulations that have been proposed in the literature, thus contributing to a unified understanding of the different formulations for this problem. We also present a novel solution approach for solving our formulation at scale, based on Benders decomposition, which exploits the surprising fact that Benders cuts for both the relaxation and the integer problem can be generated in a computationally efficient manner. We demonstrate the value of our formulation and Benders decomposition approach through two sets of experiments. In the first, we use synthetic instances to show that our formulation is computationally tractable and can be solved an order of magnitude faster for small- to medium-scale instances than the alternate, previously proposed formulations. In the second, we consider a previously studied product line design instance based on a real conjoint data set, involving over 3,000 candidate products and over 300 respondents. We show that this problem, which required a week of computation time to solve in prior work, is solved by our approach to full optimality in approximately 10 minutes.