Informs Annual Meeting 2017

WB01

INFORMS Houston – 2017

Wednesday, 10:30AM - 12:00PM

2 - Stochastic Dominance and Value of Information N. Onur Bakir, Associate Professor, Istanbul Kemerburgaz University, Istanbul Kemerburgaz Üniversitesi, Mahmutbey Dilmenler Caddesi, No:26,, Istanbul, 34217, Turkey, onur.bakir@kemerburgaz.edu.tr In this study we analyze the information acquisition problem in a multi-action setting where a decision maker chooses among multiple lotteries. In particular, we comparatively evaluate information on two lotteries that are related to each other in the sense of restricted stochastic dominance. We show in general that drawing conclusions on ranking of information for a risk averse decision maker is possible for specifically structured information alternatives. 3 - A Framework for Dealing with Baseline Uncertainty: An Analysis of China’s Carbon Intensity Target Anastasis Giannousakis, Potsdam Institute for Climate Impact Research, Telegraphenberg A.31, Potsdam, 14473, Germany, giannou@pik-potsdam.de, Elmar Kriegler, Lavinia Baumstark Intensity targets are being used as alternatives to quantity targets in the presence of baseline uncertainty. Currently, nearly one third of global emissions are under indexed control. We propose a framework to assess to what extent this is a meaningful alternative. Analyzing the Chinese intensity target with respect to situations were in multi-stage policy design either choice of target can be implemented we highlight the importance of explicitly accounting for uncertainty in economic growth, as it can lead to substantial amounts of hot-air or unexpected costs. We also identify what type of structural transformations could be more economic and how target updating creates risk of stranded assets. 4 - A New Correlation Coefficient for Aggregating Incomplete Rankings Equitably Adolfo R. Escobedo, Arizono State University, Tempe, AZ, 77802, United States, adres@asu.edu The consensus ranking problem is central to group decision-making. It involves finding an ordinal evaluation which minimizes the collective disagreement relative to a set of individual preferences over a set of competing objects; two common examples are corporate project selection and academic paper competitions. Although different measures for quantifying agreement between rankings can be employed, those founded on axiomatic distances are regarded as the most suitable due to their intuitive appeal and social choice-related axiomatic properties. This work introduces a ranking correlation coefficient founded on the Kendall tau distance metric, and it establishes its equivalence to an axiomatic ranking distance designed to handle a realistic variety of ranking formats including those containing non-strict (i.e., with-ties) and incomplete (i.e., null) preferences. Moreover, it demonstrates that alternative ranking correlation coefficients inadvertently introduce systemic biases when considering the same variety of preferences, thus rendering them inadequate for aggregating rankings in the general case. The efficacy of the presented ranking correlation coefficient to solve the consensus ranking problem and to provide alternative optimal solutions is illustrated via computational results of a new combinatorial branch-and-bound algorithm.

WB01

310A Statistics and Decision Analysis Sponsored: Decision Analysis Sponsored Session Chair: Huaiyang Zhong, Stanford University, 119 Quillen Ct, Stanford, CA, 94305, United States, hzhong34@stanford.edu 1 - Forecast Combining for Multivariate Probability Distributions Xiaochun Meng, PhD Student, Saïd Business School, University of Oxford, Park End St, Oxford, OX1 1HP, United Kingdom, xiaochun.meng@sbs.ox.ac.uk, James W. Taylor When faced with several competing forecasts, accuracy is very often improved by forming a combination of the individual forecasts. In addition to theoretical and empirical results regarding the combination of point forecasts, recent research supports the use of combining for probabilistic forecasting. For marginal distributions, it has been shown that combining quantiles is preferable to combining probabilities. We investigate whether this is also the case for joint distributions. Our empirical results concern forecasts of the joint distribution of the daily maximum and minimum temperature, which relates to the probabilistic forecasting of extreme weather events, such as heat waves. 2 - Sample Size Determination for Split Testing Based on Expected Cumulative Regret Daniel Frey, Professor, Massachusetts Institute of Technology, 77 Massachusettes Avenue, Room 3-449D, Cambridge, MA, 02139, United States, danfrey@mit.edu, Nandan Sudarsanam, Balaji Pitchai Kannu, Ravindran Balaraman Our work presents theoretical results for determining optimal sample size for A/B tests in the online setting. Unlike traditional criteria of Type 1 and 2 errors used in sample size determination, the online setting imposes the additional cost of poor performance incurred during the testing phase. We model the theoretical means of the tested alternatives, as well as the noise in the system across a range of probability distributions. In the case which explores Gaussian distributions (for both the means as well as noise), our solution for the sample size simplifies to a function of the trial horizon and a ratio of the standard deviations of the distributions. 3 - Quantile Markov Decision Processes Huaiyang Zhong, Stanford University, 119 Quillen Ct, Apt 217A, Stanford, CA, 94305, United States, hzhong34@stanford.edu, Xiaocheng Li, Margaret L. Brandeau Sequential decision making is an important task widely appeared in the fields of operations research, management science, artificial intelligence and stochastic control. In this work, we novelly introduce the framework of Quantile Markov Decision Process (QMDP) which enables us to maximize quantiles of the cumulative reward in the Markov decision process. We provide an efficient dynamic programming algorithm to solve QMDP with analysis from both theoretical and algorithmic perspectives. 310B Preferences and Risk Modeling Sponsored: Decision Analysis Sponsored Session Chair: Adolfo R. Escobedo, Arizono State University, ASU, Tempe, AZ, 77802, United States, adres@asu.edu 1 - Downside Risk Aversion and Cumulative Prospect Theory Qiulin Yang, Lancaster University, Lancaster, United Kingdom, q.yang5@lancaster.ac.uk, Zhan Pang, James Huang The increasingly popular Cumulative Prospect Theory (CPT) has challenged the Expected Utility Theory (EUT) in describing individuals’ choices of risky prospects. Our paper attempts to bridge CPT and EUT by providing a choice-theoretic characterization of the tradeoff between overall risk (variance) and downside risk for downside risk averse decision makers in the CPT paradigm. We also discuss the intrinsic relationships between the defined downside risk change and the familiar skewness measures. Moreover, we show that our analysis for S-shaped value functions can be readily extended to reverse S-shaped value functions. WB02

WB03A Grand Ballroom A Machine Learning & Causality Sponsored: Applied Probability Sponsored Session

Chair: Nathan Kallus,Massachusetts Institute of Technology, Cambridge, MA, 02139, United States, kallus@mit.edu 1 - Balanced Policy Evaluation and Learning Nathan Kallus, Cornell University, New York, NY, 02139, United States, kallus@cornell.edu Policy learning from observational data aims to transform electronic health records to personalized treatment regimes, transactional records to personalized pricing strategies, and click-streams to personalized advertising campaigns. However, existing inverse propensity weighting and doubly robust methods rely on a problematic plug-in approaches that leads to dealing with division by near zeros, data waste, and awkward two-stage procedures. I present a new approach based on optimal balance and demonstrate how it outperforms existing approaches and explain why, as supported by theory characterizing the method and regret bounds.

472