2016 INFORMS Annual Meeting Program

WD40

INFORMS Nashville – 2016

WD38 206A-MCC General Session II Contributed Session

3 - How Being Distributionally Robust Can Improve Learning In High Dimensions? Karthyek R Murthy, Columbia University, New York, NY, United States, karthyek@gmail.com, Jose Blanchet, Yang Kang In learning problems where the number of training samples is smaller than the ambient dimension, usual empirical risk minimisation may not be enough to find the best fit. We introduce RWPI, a novel learning methodology that is aimed at enhancing out-of-sample performance in such settings. By casting the learning problem as an optimization problem in the presence of model uncertainty, we recover a wide range of regularisation procedures (such as generalized Lasso, SVM) as particular cases. Further, an asymptotic analysis of a suitably defined profile function allows to optimally select the regularisation parameter. We shall discuss this optimality in the context of generalized Lasso. 4 - Learning And Hierarchies In Service Systems Michail Markakis, Universitat Pompeu Fabra, Barcelona, Spain, mihalis.markakis@upf.edu We consider a service systems with servers that have different capabilities and tasks whose types are ex-ante unknown. Information about a task’s type can only be obtained while serving it. We show that the system’s stability region depends on the entire distributions of service times, and that heavier tails cause greater performance loss. We also consider endogenizing the servers’ capabilities, and find that optimal designs have a “hierarchical” structure: all tasks are initially routed to the least skilled servers and progressively move to more skilled ones, if necessary. Comparative statics show that uncertainty in task types leads to higher training costs and less specialized server pools. WD40 207B-MCC Applied Probability and Machine Learning III Sponsored: Applied Probability Sponsored Session Chair: Daniel Russo, Northwestern University, 2001 Sheridan Road, Evanston, IL, 60208-2009, United States, dan.joseph.russo@gmail.com 1 - Collaborative Filtering With Low Regret Guy Bresler, MIT, Cambridge, MA, United States, guy@mit.edu, Devavrat Shah, Luis Voloch Empirical evidence suggests that item-item collaborative filtering (CF) works well in practice. Motivated to understand this, we provide a framework to design and analyze recommendation algorithms. The setup amounts to online binary matrix completion, where at each time a user requests a recommendation and the algorithm chooses an entry to reveal in the user’s row. The goal is to maximize the number of +1 entries revealed at any time. We analyze an item-item CF algorithm that can achieve fundamentally better performance as compared to user-user CF. Good “cold-start” performance is achieved by quickly making good recommendations to new users about whom there is little information. 2 - Predicting The Unseen Mutations Provides A Roadmap For Precision Medicine. James Zou, Stanford University, Palo Alto, CA, 02139, United States, jamesyzou@gmail.com A fundamental question in genomics is to estimate the frequency distribution of all the genetic variants in a population. This is a challenging task because we have sequenced the genomes of relatively few individuals, and most existing mutations are not observed in our samples. We give a non-parametric algorithm to estimate the frequency distribution of all the variants, including the ones that not seen in the sequenced individuals. We prove that also algorithm has strong finite-sample convergence guarantees, and applied it to one of the largest human sequencing data. Our estimates provide a roadmap for the discovery rate of large sequencing efforts, including the Precision Medicine Initiative. 3 - Causal Inference With Random Forests Stefan Wager, Stanford University, Stanford, CA, United States, swager@stanford.edu Many scientific and engineering challenges, ranging from personalized medicine to customized marketing recommendations, require an understanding of treatment heterogeneity. We develop a non-parametric causal forest for estimating heterogeneous treatment effects that extends Breiman’s widely used random forest algorithm. Given a potential outcomes framework with unconfoundedness, we show that causal forests are pointwise consistent for the true treatment effect, and have an asymptotically Gaussian and centered sampling distribution. We also propose a practical estimator for the asymptotic variance of causal forests.

Chair: Gang Wang, University of Massachusetts Dartmouth, 285 Old Westport Rd, Room 214, North Dartmouth, MA, 2747, United States, gwang1@umassd.edu 1 - Pricing Decision Model For New, Upgraded And Remanufactured Short-life Cycle Products Che-Wei Yeh, PhD Student, National Taiwan University, Floor 9, No.1, Sec. 4, Roosevelt Road, Taipei, Taiwan, d99741008@ntu.edu.tw Despite remanufacturing short life cycle products is rewarding economically as well as environmentally, very little is known about modeling upgraded decisions for products with short life cycle. In this paper, we develop a closed loop supply chain model that optimizes the price for new, upgraded and remanufactured products where demands are time-dependent and price sensitive. Using numerical analysis, the findings give reasonable results and have important implications for the impact of demand’s speed of change to the optimal prices. 2 - Effects Of The Adopting Distillers Grain In The Feed Ration For Swine Industry In Argentina Maria Celeste De Matteis, Graduate Assistant, University of Tennessee, Knoxville, TN, United States, mdematt1@vols.utk.edu, Tun-Hsiang Edward Yu Driven by the Biofuels Law enacted in 2006, the production of corn-based ethanol in Argentina has surged over the past decade. Distillers grain, a co- product of corn ethanol, is expected to become an important feedstuff in Argentine livestock feed ration because of its high content of protein and other nutrition. Our study aims to evaluate the potential impact of adopting this emerging feedstuff in the feed ration for Argentine hog industry. A multi- objective model will be developed incorporating both feedstuff cost and animal performance in the decision criteria. 3 - Integrated Operations Scheduling Under Different Penalty Terms Gang Wang, University of Massachusetts Dartmouth, 285 Old Westport Rd, Room 214, North Dartmouth, MA, 02747, United States, gwang1@umassd.edu This paper studies an integrated operations scheduling problem under service level contracts over a capacitated supply chain and considers three different scheduling sub-problems in terms of the types of service level: 1) The first sub- problem takes into account no specified service level (e.g., one-time transaction in spot contracts); 2) the second is regarding single service level contracts, where penalty function is convex; and 3) the third deals with multiple service level contracts. WD39 207A-MCC Learning and Model Uncertainty in Stochastic Systems Sponsored: Applied Probability Sponsored Session Chair: Yuan Zhong, Columbia University, New York, NY, 10025, United States, yz2561@columbia.edu 1 - Staffing Service Systems With Distributional Uncertainty John Hasenbein, University of Texas-Austin, jhas@mail.utexas.edu, Ying Chen We examine the problem of staffing service systems in which either the exact arrival rate or even the arrival rate distribution is unknown. The decision maker’s goal is to minimize staffing costs while satisfying quality-of-service constraints on the probability that a customer is delayed. We use bounds related to the Halfin- Whitt approximation and prove asymptotic optimality of the proposed methods. 2 - Ambiguous Partially Observable Markov Decision Processes Soroush Saghafian, Harvard University, Cambridge, MA, United States, soroush_saghafian@hks.harvard.edu We present a generalization of Partially Observable Markov Decision Processes (POMDPs) termed Ambiguous POMDP (APOMDP), which allows the decision maker to take into account inevitable model ambiguities. We establish various structural results, and discuss new opportunities for superior decision-making in applications such as machine replacement, medical decision-making, inventory control, revenue management, optimal search, bandit problems, and dynamic principal-agent settings.

467