INFORMS 2021 Program Book

INFORMS Anaheim 2021

TE29

decided at the end of the sales horizon due to the flexibility in choosing the number of coaches used on a train. The flexibility in varying capacity and a unique structure of the cost function distinguish this problem from the traditional RM pricing problems. Similar to the dynamic-pricing problems in RM literature, this problem is not tractable in general. Therefore, we obtain an easily implementable static policy and show that this policy is asymptotically optimal when the demand and capacity are proportionally scaled. We also obtain a dynamic policy which converges to optimality, faster than the static policy, as the scaling factor is increased. TE32 CC Room 208B In Person: Ridesharing General Session Chair: Peter Frazier, Cornell University, Ithaca, NY, 14853, United States Co-Chair: Dmitry Mitrofanov, Boston College, New York, NY, 10012- 1106, United States 1 - Pricing Fast and Slow Daniel Freund, MIT, Cambridge, MA, 02139-4165, United States, Garrett J. van Ryzin Ride hailing platforms update prices dynamically to efficiently balance supply and demand. But rapidly changing prices create incentives for riders to wait for high prices to drop. When supply builds up and prices do eventually drop, these patient customers may request en masse, causing a sharp drop in supply that triggers the pricing algorithm to increase prices. We present a simple fluid model that shows how dynamic pricing inherently creates such oscillations in supply and prices when riders are patient and strategic. Moreover, we show that these oscillations in supply levels are inherently inefficient due to the convexity of pickup times as a function of “open” (dispatchable) supply. We then show that by changing the service model to allow riders to enter a formal queue for low prices Yash Kanoria, Columbia Business School, Columbia Business School Uris Hall, New York, NY, 10027-6945, United States We consider demand and supply which arise i.i.d. uniformly in the unit hypercube [0,1]^d in d dimensions, and need to be matched with each other while minimizing the expected average distance between matched pairs (the “cost”). We characterize the achievable cost in three models as a function of the dimension d and the amount of excess supply (M or m): (i) Static matching of N demand units with N+M supply units. (ii) A semi-dynamic model where N+M supply units are present beforehand and N demand units arrive sequentially and must be matched immediately. (iii) A fully dynamic model where there are always m supply units present in the system, one supply and one demand unit arrive in each period, and the demand must be matched immediately. We show that cost nearly as small as the distance to the nearest neighbor is achievable in all cases except models (i) and (ii) for d=1 and M = o(N). 3 - Lyft and Uber IPOs: Before and After Dmitry Mitrofanov, Boston College, Chestnut Hill, MA, 02467, United States, Maxime Cohen The year 2019 witnessed two unicorn IPOs from ride-hailing platforms: Lyft filed its IPO on March 1 at a $24.3 billion valuation, and Uber filed its IPO on April 11 at a $82.4 billion valuation. Did these platforms adjust their operational decisions in anticipation of their IPOs? To answer this question, we use a comprehensive panel dataset with more than 13 million rides completed by more than 250,000 consumers between January 2018 and July 2019. We treat each IPO filing day as a natural experiment and examine how these two events have affected the operational strategies of Lyft and Uber, performance metrics, and consumers. The richness of our dataset allows us to account for various sources of heterogeneity including market penetration, loyalty, customers’ past riding frequency and sharing propensity, riders’ deal-seeking behavior, and tip amount. this inefficiency can be overcome. 2 - Dynamic Spatial Marching

TE29 CC Room 207C In Person: Frontier of Optimization and Machine Learning General Session

Chair: Yuri Fonseca, New York, NY, 10027, United States 1 - Offline and Online Learning from Optimal Actions Yuri Fonseca, Columbia University, New York, NY, 10027, United States, Omar Besbes, Ilan Lobel We study the offline and online problem of contextual optimization where instead of observing the loss, we observe the optimal action an oracle with full knowledge would have taken. At each period, the decision-maker has access to a new set of feasible actions to select from and to a new contextual function that affects that period’s loss function. In the offline setting, the decision-maker has already collected information from multiple periods. We aim to minimize regret, which is defined as the difference between our losses and the ones incurred by an all-knowing oracle. Through our offline analysis, we tightly connect the type of performance that can be achieved as a function of the underlying geometry of the information induced by offline data. For the online setting, we leverage this tight link to optimize regret. 2 - Finite Sample Analysis of Minimax Offline Reinforcement Learning: Completeness, Fast Rates and First-order Efficiency Masatoshi Uehara We offer a theoretical characterization of off-policy evaluation (OPE) in reinforcement learning using function approximation for marginal importance weights and q-functions when these are estimated using recent minimax methods. Under various combinations of realizability and completeness assumptions, we show that the minimax approach enables us to achieve a fast rate of convergence for weights and quality functions, characterized by the critical inequality \citep{bartlett2005}. Based on this result, we analyze convergence rates for OPE. In particular, we introduce novel alternative completeness conditions under which OPE is feasible and we present the first finite-sample result with first-order efficiency in non-tabular environments, i.e., having the minimal coefficient in the leading term. In Person: Featured Session: Reflection on the Opportunities and Challenges of COVID-19 Panel Session Chair: Abdallah A Chehade, University of Michigan-Dearborn, Dearborn, MI, 48128-2406, United States Chair: Xiao Liu, University of Arkansas, Fayetteville, AR, 72701, United States 1 - Moderator Abdallah A. Chehade, University of Michigan-Dearborn, HPEC, Dearborn, MI, 48128-2406, United States Distinguished panelists from the National Science Foundation and Academia will discuss the impacts of COVID-19 on academic research, teaching, and service. The discussion will be focused on the challenges, practices, and opportunities during and post COVID-19. 2 - Panelist Kamran Paynabar, ISyE Georgia Tech, Georgia Tech, H. Milton Stewart School Of Isye, Atlanta, GA, 30332-0205, United States TE31 CC Room 208A In Person: Learning Algorithms in Revenue Management General Session Chair: Anyan Qi, The University of Texas at Dallas, Richardson, TX, 75080-3021, United States 1 - A Joint Pricing and Capacity Decision Problem in Railways Seetharama Chandrasekhar Manchiraju, University of Texas- Dallas, Naveen Jindal School of Man., Richardson, TX, TE30 CC Room 207D

75080-3021, United States, Milind Dawande, Ganesh Janakiraman, Arvind Raghunathan

We study a joint pricing and capacity decision problem in the railway industry. Unlike in industries such as aviation, capacity in the railway industry can be

146