Informs Annual Meeting 2017

TA01

INFORMS Houston – 2017

MD82

2 - Statistical Optimality and Model Assessment for SP in Learning Enabled Optimization (LEO) Yunxiao Deng, PhD Student, University of Southern California, Los Angeles, CA, 90089, United States, yunxiaod@usc.edu For Learning Enabled Optimization problems, we can obtain solutions with statistical guarantees using concepts of statistical optimality. Model assessment and selection are also critical parts of the modeling process. We present a collection of hypothesis tests, metrics and comparisons between alternative LEO models to assess solution quality. 3 - Ensemble Bayesian Optimization for Partially Controllable Sequential Information Processes Yingfei Wang, Princeton University, Department of Computer Science,, Princeton University, 35 Olden Street,, Princeton, NJ, 08540-5233, United States, yingfei@princeton.edu, Warren B.Powell An important element in sequential decision making problems is stochastic models of the environment and proper statistical models and inferences to represent our changing beliefs about the environment as new information is collected. We design an ensemble optimal learning method to respond quickly and robustly to complex data streams. In our ensemble system, multiple models, such as classifiers, are strategically generated and combined to minimize the incorrect selection of a particularly poorly performing statistical model. Similar to the idea of online boosting, we use Bayesian learning with expert advice as the belief model, aiming to spend the limited measurement budget more wisely. 4 - Learning Rare-event Probabilities and Applications to Automated Vehicle Testing Zhiyuan Huang, University of Michigan, Ann Arbor, MI, 48109, United States, zhyhuang@umich.edu, Henry Lam, Ding Zhao This talk discusses the estimation of rare-event probabilities where learning takes place in the measurement of the hitting set of interest. Because of the rare-event nature, the estimation requires the use of variance reduction techniques that are sequentially optimized. This problem arises from the safety testing of automated vehicles using simulation-based estimation of critical events such as crashes or conflicts, where the vehicles’ behaviors are analytically intractable or not fully observable.

382B Optimizing Network Robustness under Uncertainty Sponsored: Optimization, Optimization Under Uncertainty Sponsored Session Chair: Tanveer Hossain Bhuiyan, tb2038@msstate.edu 1 - Chance Constrained Distributional Robust Optimization for Capacity Expansion of Wind Generation Wei Wang, Doctoral Student, University of Pittsburgh, 1025 Benedum Hall, Pittsburgh, PA, 15261, United States, w.wei@pitt.edu, Bo Zeng To handle uncertainties in wind energy generation in capacity expansion, a chance-constrained distributional robust optimization model is developed and solved. Numerical results will be reported. 2 - Algorithm for a Stochastic Bilevel Optimization Chaosheng Dong, University of Pittsburgh, 4200 Fifth Avenue, Pittsburgh, PA, 15260, United States, chaosheng@pitt.edu, Bo Zeng For a bilevel optimization problem with multiple lower level problem, we analyze its property and develop a distributed computational algorithm. Numerical results will be presented. 3 - Network Design and Facility Protection when the Effect of Protection is Uncertain Tanveer Hossain Bhuiyan, Mississippi State University, P.O. Box 5227, Room 321, Starkville, MS, 39762, United States, tb2038@msstate.edu, Hugh Medal We study a facility location and network design problem that involves protecting facilities subject to random disruptions where the protection is imperfect, multi- level, and the effect of disruption is imperfect. The goal of our study is to optimally allocate protection resources to the facilities, and construct links in the network to minimize the expected total transportation cost. We model the problem as a two-stage stochastic programming with decision dependent uncertainty where the post-disruption capacity states of the facilities depends probabilistically on the resource allocation decision and the disruption intensity. We implement an L-shaped algorithm to solve the model. 4 - Two-stage Robust Optimization for Single-commodity Network Design with Potential Edge Failures Logan R. Matthews, Princeton University, Princeton, NJ, United States, loganm@princeton.edu, Yannis G. Kevrekidis, Chrysanthos E. Gounaris The single-commodity network design problem with potential edge failures is formalized as a two-stage robust optimization problem with an uncertainty set of binary random variables representing uncertain edge failures. A column-and- constraint generation algorithm is tailored to this application and implemented to gain minimum cost networks which remain feasible for a given number of simultaneous edge failures. Problems adapted from the Survivable Network Design Library are solved to measure the cost increases and computational efforts required due to various levels of robustness. 382C Stochastic Optimization with Learning Sponsored: Optimization, Optimization Under Uncertainty Sponsored Session Chair: Daniel Jiang, University of Pittsburgh, University of Pittsburgh, Pittsburgh, PA, 15261, United States, drjiang@pitt.edu 1 - Optimal Online Learning for Nonlinear Belief Models with Physical State Information Weidong Han, Princeton University, 35 Olden Street, Sherrerd Hall, Princeton, NJ, 08544, United States, whan@princeton.edu, Warren Powell We consider an optimal learning problem with nonlinear belief models and physical state information such as a finite budget or resource. We assume the parameter vector in the belief model follows a discrete prior distribution. We formulate the problem into a dynamic program, and propose a lookahead policy called the OC-correction that is shown to be optimal for two-stage problems. We prove asymptotic convergence properties of the proposed policy for infinite- horizon problems. Adopting this framework, we find an approach to the real-time bidding problem for online advertising. MD83

MD84 General Assembly A Markov Lecture Invited Session

1 - Methods for Model Approximation and Optimization in the Presence of Model Uncertainty using Information Divergences Guodong Pang, Penn State University, University Park, PA Tuesday, 7:30 - 9:00AM

TA01

310A Modeling, Decisions, and Sensitivity Sponsored: Decision Analysis Sponsored Session Chair: Emanuele Borgonovo, Bocconi University, Milano, 20136, Italy, emanueleborgonovo@unibocconi.it Co-Chair: Alessandra Cillo, Bocconi University, Milan, 20136, Italy, alessandra.cillo@unibocconi.it 1 - Multinomial Logit Processes and Preference Discovery Fabio Maccheroni, Università Bocconi, Milan, Italy, fabio.maccheroni@unibocconi.it, Simone Cerreia-Vioglio, Massimo Marinacci, Aldo Rustichini We study and axiomatically characterize the dependence of choice probabilities of an agent on the time available to decide by means of a Multinomial Logit Model with time independent (systematic component of the) utility and time dependent scale parameter. This is the most widely used model of preference discovery in all fields of decision making from Quantal Response Equilibrium theory to Discrete Choice Analysis, from Psychophysics to Combinatorial Optimization. Our axiomatic characterization permits, on the one hand, to understand its conceptual underpinnings as a theory of agents behavior, on the other to empirically test its descriptive validity.

267