INFORMS 2021 Program Book

INFORMS Anaheim 2021

TD36

2 - Quality Learning in a Dynamic Mutual Data Exchange Model Ebru Kasikaralar, University of Chicago, Chicago, IL, United States, John R. Birge Advances in storing and processing big data have transformed how digital platforms learn about product quality, user preferences and make pricing decisions accordingly. Due to the increase in data privacy concerns, policies restricting access to consumers’ data are expected to become more prevalent. However, personal data is essential in generating successful matching algorithms and enhancing the customers’ purchase experience. Hence, we introduce a model where the buyers and a firm directly interact with each other in two different markets: product and data market. There is a costly dynamic data exchange between the firm and the consumers, where the firm offers incentives to buyers to sell data. The firm uses acquired data from consumers to learn the underlying product quality, and consumers use the acquired data from the firm to make strategic purchase decisions. 3 - Data-driven Newsvendor: Algorithms and Optimal Performance Omar Mouchtaki, Columbia University, New York, NY, United States, Omar Besbes We study the classical newsvendor problem in which the decision-maker must make decisions to trade-off underage and overage costs. In contrast to the typical setting, we assume that the decision-maker does not know the underlying distribution driving uncertainty but has only access to past data drawn from the underlying distribution (e.g., past demand). In turn, the key question is how to map existing data to an optimal decision. We evaluate the performance of any algorithm through its worst-case relative expected regret, compared to an oracle with knowledge of the distribution. We provide the first finite sample exact analysis of the classical Sample Average Approximation (SAA) algorithm for this class of problems across all data sizes. We further derive an optimal algorithm and its associated performance. It yields significant improvements over SAA for small data sizes. 4 - Deep Learning For Visual Advertising on Digital Platforms Yuexing Li, Duke University, Durham, NC, 27703-6548, United States, N. Bora Keskin, Shaoxuan Liu, Jing-Sheng Jeannette Song We consider a digital platform that aims to crop and display N images to its customers to help with their purchasing decisions. For each image, the platform chooses a cropping window and observes the resulting conversions, i.e., the customer purchasing decisions. The platform does not know how cropped images influence conversions. We design a neural network policy that dynamically learns this relationship and adjusts images to maximize conversion. We derive a theoretical performance guarantee proving the asymptotic optimality of our policy. Using real-life data from a large online travel platform, we show that our policy achieves considerable improvement over the incumbent policy of the platform. The results also reveal that our policy exhibits good performance even if the functional relationship between images and conversion is misspecified. Chair: Bingqing Liu, New York University, New York, NY, United States 1 - Ridesharing Morning Commute in Monocentric City Networks an Equilibrium Model and the Analytical Solutions Rui Ma, University of Alabama in Hunstville, Owens Cross Roads, AL, 35763, United States The ridesharing morning commute traffic in a many-to-one network where commuters from different origins commute to the central business district (CBD) is studied. It is found that the common parking disutility connects departure-time choice behavior and traffic flow patterns on all corridors.Seemingly counter- intuitive, a demand paradox and a corridor expansion paradox are found, which have significant implications for both urban traffic management and infrastructure planning for concentric cities. 2 - Ridesharing and Fleet Sizing for On-demand Multimodal Transit Systems Ramon Auad, ISyE Georgia Tech, Atlanta, GA, 30318-5499, United States, Pascal Van Hentenryck This work considers the design of On-Demand Multimodal Transit Systems (ODMTS) that combine fixed bus/rail routes between transit hubs with on- demand shuttles to serve the first/last miles to/from the hubs. The design problem aims at finding a network design for the fixed routes to allow a set of riders to travel from their origins to their destinations while minimizing the sum of the travel costs, the bus operating costs, and rider travel times. Using MIP models, the paper generalizes prior work by including ridesharing in the shuttle rides and proposes a novel fleet-sizing algorithm for determining the number of shuttles needed to meet the performance metrics of the ODMTS design. The methodological contributions are evaluated on a real case study in Michigan to illustrate the potential of ridesharing for ODMTS. TD36 CC Room 210B In Person: Emerging Themes in Urban Transportation Planning General Session

TD32 CC Room 208B In Person: Innovative Business Models and Platforms General Session Chair: Longyuan Du, University of San Francisco, San Francisco, CA, 94117-1080, United States 1 - The Important Role of Time Limits When Consumers Choose Their Time in Service Pnina Feldman, Boston University, Questrom School of Business, 595 Commonwealth Aven, Boston, MA, 2215, United States, Ella Segev We examine ways to manage congestion in services when customers choose their service time. We find that time limits are very attractive levers. They are optimal for revenue and social welfare maximization and are nearly optimal to maximize consumer surplus. 2 - Sales Effort Management under All-or-nothing Constraint Longyuan Du, University of San Francisco, San Francisco, CA, 94117-1080, United States, Ming Hu, Jiahua Wu We consider a sales effort management problem under an all-or-nothing constraint. The seller will receive no bonus/revenue if the sales volume fails to reach a predetermined target at the end of the sales horizon. Throughout the sales horizon, the sales process can be moderated by the seller through costly effort. We show that the optimal sales rate is non-monotonic with respect to the remaining time or the outstanding sales required to reach the target. We then study easy-to- compute heuristics. We show that when the profit-maximizing rate is lower than the target rate, the performance loss of any static heuristic is of an order greater than the square root of the scale parameter. To address the poor performance of the static heuristic, we propose a modified resolving heuristic and show that it is asymptotically optimal and achieves a logarithmic performance loss. Management General Session Chair: Yuexing Li, Duke University, Durham, NC, 27703-6548, United States 1 - Provably Optimal Reinforcement Learning for Online Inventory Models With Cyclic Demands Xiao-Yue Gong, Massachusetts Institute of Technology, Cambridge, MA, 02139-4301, United States, David Simchi-Levi Motivated by a long-standing gap between inventory theory and operations practice, we study online inventory models with unknown cyclic stochastic demands. We design efficient reinforcement learning algorithms with provable theoretical guarantees that cater to the specific structures of inventory management problems. We apply a standard performance measure in online learning literature, regret, defined as the difference between the expected total cost of a policy and that of the optimal policy with full knowledge of the demand distributions.This paper first introduces an easier family of models—episodic models—where inventory is discarded at the end of every cycle. Our policies achieve ~O(\sqrt{T}) regret for both the episodic lost-sales model with zero lead time, and the episodic multi-product backlogging model with lead times, fixed joint-ordering costs and order limits. The regret bounds of our policies match the regret lower bounds that we prove for these models. Next, we build upon these policies for episodic models to devise meta algorithms for the more difficult non- discarding models with cyclic demands. Our policies achieve ~O(\sqrt{T}) regret for both the lost-sales model with zero lead time and the backlogging model with zero lead time, again matching the regret lower bounds for these models. We achieve ~O(T^{5/6}) regret for the multi-product non-discarding backlogging model with lead times, fixed joint-ordering costs and order limits. Our policies allow the input of expert advice to further improve their performance in real applications. Importantly, some of our algorithms apply more generally to a variety of operations research problems beyond inventory management. TD33 CC Room 209A In Person: Data-driven Learning in Revenue

135