2016 INFORMS Annual Meeting Program

TD47

INFORMS Nashville – 2016

TD45 209A-MCC Simulation Optimization and Ranking and Selection Sponsored: Simulation Sponsored Session Chair: Demet Batur, University of Nebraska-Lincoln, CBA 209, Lincoln, NE, 68588, United States, dbatur@unl.edu 1 - Optimization-based Learning Of Simulation Model Discrepancy Henry Lam, University of Michigan, Ann Arbor, MI, United States, khlam@umich.edu, Matthew Plumlee The vast majority of stochastic simulation models are imperfect in that they fail to fully emulate the entirety of real dynamics. Despite this, these imperfect models are still useful in practice, so long as one knows how the model is inexact. We propose a method to learn the amount of the model inexactness using data collected from the system of interest. Our approach relies on a Bayesian framework that addresses the requirements for estimation of probability measures that are ubiquitous in stochastic simulation, and an embedded optimization to enhance the involved computational efficiency. 2 - Quantile-based Ranking And Selection In A Bayesian Framework Yijie Peng, Fudan University, pengy10@fudan.edu.cn, Chun-Hung Chen, Michael Fu, Jian-Qiang Hu, Ilya O Ryzhov We propose two quantile-based ranking and selection schemes in a Bayesian framework, i.e. myopic allocation policy (MAP) and optimal computing budget allocation (OCBA). MAP has a superior small sample performance, while OCBA shows a desirable asymptotic behavior. As a result, a switching strategy that that switches from MAP to OCBA is provided to achieve balanced performances in both small sample and large sample scenarios. 3 - Multi-information Source Optimization With Applications Matthias Poloczek, Cornell University, poloczek@cornell.edu, Jialei Wang, Peter Frazier We consider Bayesian optimization of an expensive-to-evaluate black-box function, where we also have access to cheaper approximations of the objective that are typically subject to varying unknown bias. Our novel algorithm rigorously treats the involved uncertainties and uses the Knowledge Gradient to maximize the predicted benefit per unit cost. We discuss applications and demonstrate that the method consistently outperforms other state-of-the-art techniques, finding designs of considerably higher objective value at lower cost. 4 - Tractable Dynamic Sampling Strategies For Quantile-based Ordinal Optimization Dongwook Shin, Columbia Business School, dshin17@gsb.columbia.edu, Mark Nathan Broadie, Assaf Zeevi Given a certain number of stochastic systems, the goal of our problem is to dynamically allocate a finite sampling budget to minimize the probability of falsely selecting non-best systems, where the selection is based on quantiles of their performances. The key aspect is that the objective depends on underlying probability distributions that are unknown. To formulate this problem in a tractable form, we introduce a function closely associated with the aforementioned objective. To derive sampling policies that are practically implementable, we suggest a policy that combines sequential estimation and myopic optimization, as well as certain variants of this policy for finite-time improvement. TD46 209B-MCC Empirical Research on Pricing and Revenue Management Sponsored: Revenue Management & Pricing Sponsored Session Chair: Pnina Feldman, University of California-Berkeley, Berkeley, CA, United States, feldman@haas.berkeley.edu Co-Chair: Necati Tereyagoglu, Scheller College of Buss - Georgia Institute of Technology, Atlanta, GA, United States, necati.tereyagoglu@scheller.gatech.edu 1 - Welfare Implications Of Congestion Pricing: Evidence from SFpark Hsin-Tien Tsai, University of California, Berkeley, 1822 Francisco

2 - Inventory Announcements And Customer Choice: Evidence From The Air Travel Industry

Katherine Ashley, University of California-Berkeley, kate_ashley@haas.berkeley.edu, Pnina Feldman, Jun Li

Does inventory announcement influence consumer decision-making in the market for airline tickets? We estimate the impact of the firm’s announcement policy on customer purchase timing and itinerary choice. In doing so, we measure the information content of inventory announcements, and analyze the extent to which customers treat these messages from the firm as cheap talk or credible information. 3 - Designing Listing Policies For Online B2b Marketplaces Wenchang Zhang, University of Maryland, COLLEGE PARK, MD, CA, United States, wzhang@rhsmith.umd.edu, Konstantinos Bimpikis, Wedad Jasmine Elmaghraby, Kenneth Moon Excess inventory amounts to $500 billion a year for big-box retailers. Much of this inventory is sold through online B2B auctions. Based on a natural experiment, we provide strong evidence that increasing the market thickness by concentrating the auction ending times to just a couple of days of the week has a significant positive effect on their final prices. We find that bidders’ monitoring cost has a large impact on their auction entry choices and outweighs the potentially negative effect of cannibalization among competing auctions. Our findings may have implications for the design of online marketplaces beyond liquidation auctions. 4 - Distribution Channel Relationships And Multimarket Competition Necati Tereyagoglu, Assistant Professor of Operations Management, Scheller College of Bussiness - Georgia Institute of Technology, 800 W Peachtree St NW, Atlanta, GA, 30308, United States, necati.tereyagoglu@scheller.gatech.edu, O. Cem Ozturk We study the role of the distribution channel relationships in determining competitive intensity when manufacturers encounter in multiple markets. We explore the manufacturers’ pricing decisions when they have asymmetric distribution channel relationships with retailers across multiple markets. Using an extensive scanner data set, we find that cross-market interdependence due to shared ties with the retailers softens competition when manufacturers have asymmetric distribution channel relationships across multiple markets. TD47 209C-MCC Cloud Computing in RM Sponsored: Revenue Management & Pricing Sponsored Session Chair: Cinar Kilcioglu, Columbia Business School, Columbia Business School, New York, NY, 10027, United States, ckilcioglu16@gsb.columbia.edu Co-Chair: Costis Maglaras, Columbia University, New York, NY, United States, c.maglaras@columbia.edu 1 - Optimal Resource Consumption via Data Driven Prophet Inequalities With An Application To Cloud Infrastructure Andrew A Li, MIT, Cambridge, MA, United States, aali@mit.edu, Muhammad J Amjad, Vivek Farias, Devavrat Shah Buyers of cloud compute resources are generally interested in completing workloads by fixed deadlines as cheaply as possible. This entails purchasing enough resources at the lowest prices possible, which is a challenge in today’s market where the largest providers all use some form of demand-driven pricing. We formulate this as a covering problem, and introduce the Data-Driven Prophet Model, which uses historical price data to interpolate between stochastic modeling and a fully adversarial model. We propose a simple, scalable threshold policy that is order-optimal and has, in a real-world implementation, completed workloads significantly cheaper than the current practice benchmark. 2 - Stochastic Optimal Control Of Time-varying Cloud Workloads Mark S Squillante, IBM Thomas J. Watson Research Center, mss@us.ibm.com, Yingdong Lu, Mayank Sharma, Bo Zhang We consider a cloud computing system modeled as a GI/GI/1 queue under workloads (arrival and service processes) that vary on one time scale and under controls (server capacity) that vary on another time scale. Taking a stochastic optimal control approach and formulating the corresponding optimal dynamic control problem as a stochastic dynamic program, we derive structural properties for the optimal dynamic control policy in general. We also derive fluid and diffusion approximations for the problem and propose analytical and computational approaches in these settings. Computational experiments demonstrate the benefits of our approach.

St., Apt 4, Berkeley, CA, 94703, United States, hsintien@berkeley.edu, Pnina Feldman, Jun Li

SFpark is a congestion pricing program for street parking implemented in San Francisco. We investigate whether consumers benefit from congestion pricing using data from this program. We build a structural model of consumer search and quantify the change in consumer welfare.

349