Informs Annual Meeting Phoenix 2018

INFORMS Phoenix – 2018

SC41

3 - Distributionally Robust Optimal Bidding in First-Price Auctions Jose Blanchet, Stanford University, Stanford, CA, United States Until recently, second price auctions overwhelmingly dominated the market making mechanics in online advertising, but there is a clear trend in which online advertising exchanges are using first-price auctions (and other types of mechanisms) for market making. Optimal bidding in these types of auctions requires highly detailed knowledge on the bidder’s competitive landscape, making it extremely challenging for Demand Side Platforms (or DSPs) to manage strong campaigns on behalf of the advertisers. So, distributionally robust optimization (DRO) is particularly well-suited in this setting. We introduce a framework based on DRO which is simple, interpretable and provides reasonable bidding policies which are well-suited for DSP use. 4 - Sample-based Optimal Pricing Amine Allouah, Columbia University,New York, NY, 10027, United States, Omar Besbes We study the problem of optimal pricing to sell one indivisible good to a buyer, when facing limited information about the willingness-to-pay distribution. In particular, we characterize optimal performance based on a single sample. n SC40 North Bldg 226B Stochastic Models and Control Sponsored: Applied Probability Sponsored Session Chair: Rami Atar, Technion, Technion, Haifa, 32000, Israel 1 - Some Notions of Minimality for Resource Sharing Networks and their Fluid Models Lukasz Kruk, Maria Curie-Sklodowska University, Lublin, Poland We consider real-time stochastic resource sharing networks with soft customer deadlines and arbitrary topology. For such systems, we introduce several notions of pathwise minimality and we discuss their relations to the Earliest-Deadline- First service discipline. We also mention some counterparts of these minimality concepts for fluid models of resource sharing networks. 2 - Large Deviations for M/M/1 Queue with Abandonment and Long-time Analysis Ruoyu Wu, University of Michigan, Ann Arbor, MI, United States, Rami Atar, Amarjit Budhiraja, Paul Dupuis We consider M/M/1 queue with abandonment according to exponential clocks. A large deviation principle is established for the queue length and total abandonment processes. We study the exponential decaying rate, and its long- time asymptotics, of the probability that there are more abandonment than the law of large numbers limit. Discussions on analyzing related risk sensitive control problems and large deviation bounds of G/G/1 queue with abandonment, using variational formula of the Renyi divergence, are also provided. 3 - A Fluid Limit for an Overloaded Multi-class Many-server Queue with General Reneging Distribution Amy R. Ward, University of Southern California, Marshall School of Business, Bridge Hall BRI 401H, Los Angeles, CA, 90089-0809, United States, Amber L. Puha We study scheduling in a many server queue with general reneging distribution (G/G/N+G) and multiple customer classes. Motivated by the goal to formulate and analyze a fluid control problem, we specify a class of admissible control policies (rules for determining when to serve a given customer class), formulate a fluid model, and characterize the invariant states. Static priority policies do not capture the entire spectrum of invariant states, and so we introduce a set of control policies, called Random Buffer Selection (RBS) that do. We prove that a suitably rescaled state descriptor for the queue operating under specified RBS policy converges to the unique fluid model solution for that RBS policy. 4 - Subgeometric Ergodicity of Levy Driven SDEs Arising from Multiclass Many Server Queues Ari Arapostathis, University of Texas at Austin, Austin, TX, USA, Guodong Pang, Nikola Sandric We study a class of multidimensional piecewise Ornstein-Uhlenbeck processes with jumps, which contains limiting diffusions arising in multiclass many-server queues subject to heavy-tailed arrivals and/or asymptotically negligible service interruptions in the Halfin-Whitt regime as special cases. We identify conditions on the parameters in the drift, the Levy measure and/or covariance function which result in subexponential and/or exponential ergodicity. We show that these assumptions are sharp, and we identify some key necessary conditions for the process to be ergodic. In addition, we show that for the queueing models described above with no abandonment, the rate of convergence is polynomial, and we provide a sharp quantitative characterization of this rate via matching upper and lower bounds.

n SC41 North Bldg 226C Preferences in Project Evaluation and Selection Sponsored: Decision Analysis Sponsored Session Chair: Enrico Diecidue, INSEAD, Fontainebleau Cedex, 77305, France 1 - Net Present Value Analysis of Projects under Expected Utility Manel Baucells, University of Virginia, Darden School of Business, 100 Darden Boulevard, Charlottesville, VA, 22903, United States, Samuel Bodily The traditional decision-analytic approach to evaluate projects is to calculate the expected utility of initial capital plus net present value. The choice of discount rate, and the convergence with the traditional finance approach, have always been a question. Our goal is to fill this gap. Under assumptions compatible with the CAPM model, we find a convenient rate to discount portfolio capital (treasuries, stocks, and the project). We also cater to practitioners, who discount the cash flows of the project, and ignore the market uncertainty. For them, we propose an adjusted discount rate that correctly compensate for the omission. 2 - Probability Dominance Enrico Diecidue, INSEAD, Boulevard De Constance, Fontainebleau Cedex, 77305, France We test whether a simple heuristic of maximizing the probability of being ahead, probability dominance (PD), affects decisions under risk. We set up head-to-head situations where all preferences of a given class (expected utility, prospect theory, or regret theory) are for one alternative yet PD favors the other. Our experiments reveal that: the majority of choices are aligned with PD in contradiction to any form of expected utility maximization, prospect theory preferences, and regret theory. We conclude that probability dominance affects choices, and should therefore be incorporated into decision making models. We quantify a lower bound on its weight in the decision-making process. 3 - Optimal Switching Between Projects with Different Profitability and Uncertainty Characteristics Tord Olsen, PhD Student, Norwegian University of Science and Technology, Alfred getz vei 3, Trondheim, 7491, Norway, Verena Hagspiel We consider a firm with an investment opportunity to introduce a new product, where the profit is modelled as a geometric Brownian motion, and the drift and volatility of the underlying process changes after investment. We find that under the assumption of a concave profit function, the conventional insight of investment under uncertainty, that an increase in uncertainty delays investments, might not hold if the characteristics of the stochastic process changes after the investment. Further, we show that a case with a boost in the profit function under constant parameters of the stochastic process can be transformed to a case of constant profit function with changes to the underlying process. 4 - Decomposition of Sensitivity Analysis for Problems with Correlated Inputs Warren Joseph Hahn, The University of Texas at Austin, McCombs School of Busines, 217 Grant Cannon Lane, Austin, TX, 78738, United States, Tianyang Wang, James Dyer The state of the art for sensitivity analysis of decision making problems with dependencies between input variables are fully probabilistic techniques which simultaneously vary all input variables. However, much of the intuition desired from sensitivity analysis is not provided with these methods because they do not isolate the marginal effects of the individual variables. In this paper, we present an approach that provides results that are identical to those from a fully- probabilistic sensitivity analysis, but which also decomposes the sensitivity to each individual input variable into its marginal dependence-constrained sensitivity and its sensitivity due to dependence with other inputs.

75