Informs Annual Meeting Phoenix 2018

INFORMS Phoenix – 2018

WD39

3 - Determining Optimal Parameters for an Expediting Policy under a Service Level Constraint Simon Hoeller, University of Cologne, Albertus-Magnus-Platz, Cologne, 50923, Germany, Raik Oezsen, Ulrich Thonemann We consider a periodic review inventory system with stochastic demand, deterministic lead times, back-ordering, and the option to move outstanding units forward in the replenishment pipeline. The objective is to minimize inventory holding and expediting costs per period subject to a minimum service level constraint. We consider a generalized base-stock policy where outstanding units are expedited when the inventory level drops below a certain threshold. We develop structural properties and present an efficient procedure to determine optimal policy parameters. In a numerical study, we show that the expediting policy offers substantial savings compared to the classical base-stock policy. 4 - An Aggregation-based ADP Approach for the Periodic Review Model with Random Yield Ulrich Thonemann, Universitat zu Koln, Wirtschafts und Sozialwissenschaftliche, Fakultat Albertus Magnus Platz, Koeln, D-50923, Germany, Michael Voelkel, Anna-Lena Sachs A manufacturer places orders periodically for products that are shipped from a supplier. Orders may get damaged with some probability, i.e. the order is subject to random yield. The manufacturer may track its orders to receive information on damages and to place additional orders. We solve this model with stochastic demand, tracking cost and random yield in all periods to optimality. We propose a novel aggregation-based approximate dynamic programming algorithm and provide solutions for larger instances for which it is not feasible to obtain optimal solutions. We analyze the effect of dynamic tracking and develop a heuristic that takes tracking costs into account to solve even larger instances. 5 - A Newsvendor Analysis with Carbon Emission Regulations Sungyong Choi, Assistant Professor, Yonsei University, College of Government and Business, 1 Yoneseidae-gil, Wonju, 26493, Korea, Republic of This paper aims to provide an optimization model for operational efficiency in individual firms considering various types of carbon emission regulations. More specifically, this study assumes that customer demand is given as a probability distribution. Under this circumstance, I formulate newsvendor models including carbon emission regulations and then derive practical implications for the policymakers in carbon emission regulations. Then, I analyze the models to provide closed-form solutions and conduct a sensitivity analysis for the impacts of model parameters on the optimal solutions through a comparative static analysis. All analytical results are reconfirmed by numerical analysis. 6 - Inventory Management under Corporate Income Tax Yixuan Xiao, City University of Hong Kong, 83 Tat Chee Avenue, Kowloon, Hong Kong, Zhan Pang Corporate income tax is a significant cost for companies and an important input into many corporate decisions. Corporations use after-tax earnings to reinvest in their core business and pay out dividends. We propose a framework to study a firm’s inventory decisions under taxation in multi-accounting periods where each accounting period consists of multiple ordering periods. We show that in the backlog model under a convex tax or in the lost-sales model under a flat-rate tax and mild conditions, a state-dependent base-stock policy is optimal. We also examine the static effect and intertemporal effect of tax on a firm’s inventory decisions. n WD37 North Bldg 225A Applied Probability in Machine Learning and Optimization I Sponsored: Applied Probability Sponsored Session Chair: Chang-Han Rhee, Centrum Wiskunde and Informatica, Amsterdam, 1018 MZ, Netherlands 1 - Approximating Data-driven Joint Chance-constrained Programs via Uncertainty Set Construction Zhiyuan Huang, University of Michigan, Ann Arbor, MI, 48109, United States, Jeff Hong, Henry Lam We discuss a statistical framework to integrate data into robust optimization, based on prediction set learning and a simple data-splitting validation scheme that achieves finite-sample statistical guarantees on the feasibility of the underlying uncertain constraints. We demonstrate several features of the framework, including a dimension-free sample size requirement for the feasibility guarantees and the capability to self-improve existing solutions in terms of both optimality and feasibility.

2 - Offline Multi-action Policy Learning: Generalization and Optimization Zhengyuan Zhou, 160 Comstock Circle - Unit 106002, Stanford, CA, 94305, United States, Susan Athey, Stefan Wager The unprecedented growth of easily accessible user-specific data has welcomed the exciting era of personalized decision making, an ubiquitous paradigm that has been revolutionizing many areas of operations research (e.g. health care, Internet advertising, product recommendation, public policies). The central problem in personalized decision making lies in learning from observational data a good policy, which provisions decisions based on each individual’s distinct set of characteristics. In this work, we study the general offline policy learning problem and provide a principled framework to learn an effective policy from data. 3 - The Nearest Neighbor Information Estimator is Adaptively Near Minimax Rate-optimal Jiantao Jiao, Stanford University, CA, United States, Weihao Gao, Yanjun Han We analyze the Kozachenko-Leonenko (KL) nearest neighbor estimator for the differential entropy. We obtain the first uniform upper bound on its performance over Holder balls on a torus without assuming any conditions on how close the density could be from zero. Accompanying a new minimax lower bound over the Holder ball, we show that the KL estimator is achieving the minimax rates up to logarithmic factors without cognizance of the smoothness parameter s of the H¿older ball for $s\in (0,2]$ and arbitrary dimension d, rendering it the first estimator that provably satisfies this property. 4 - Data Driven Distributionally Robust Optimization via Optimal Transport: Algorithms and Applications Fan Zhang, Stanford University, 450 Serra Mall, Stanford, CA, 94305, United States, Jose Blanchet, Yang Kang, Karthyek Murthy In this talk, we are going to explain why data-driven distributionally robust optimization (DRO) is an important class of stochastic optimization problems by showing that it encompasses popular estimators in machine learning. We will provide optimal algorithms under general DRO formulations. In particular, we will prove that solving a DRO problem of an affine decision model is not harder than solving its non-robust counterpart. We will also show some estimates of the optimal transport rules. Finally, we will demonstrate empirically that our proposed methodology is able to improve upon popular machine learning estimators. n WD39 North Bldg 226A Joint Session APS/Opt-Uncert: Robust and Dynamic Stochastic Optimization Sponsored: Applied Probability Sponsored Session Chair: Saumya Sinha, University of Washington, Seattle, WA, 98195, United States 1 - A Utility Uncertainty Approach to Multiattribute Bayesian Optimization Raul Astudillo, Cornell University, Ithaca, NY, 14850, United States We consider multi-attribute Bayesian optimization, where each feasible design is associated with a vector of attributes that can be evaluated via a time-consuming computer code, and each vector of attributes is assigned a utility according to a decision-maker’s implicit utility function. We propose a sampling policy that maximizes the expected utility of the design chosen by the decision-maker, where her choice is based on the policy’s sampling- based attribute vector estimates. In contrast with existing approaches for multi-attribute optimization that focus on estimating a Pareto frontier, our approach leverages prior information about the decision-maker’s preferences. 2 - A Distributionally Robust Capacity Planning Model for Optimizing Access Delay in Surgical Services under Limited Information Mohammad Zhalechian, PhD Student, University of Michigan, 1205 Beal Avenue, Ann Arbor, MI, 48109-2117, United States, Mark P. Van Oyen We develop a capacity allocation framework that reserves OR time slots for different types of surgeries (department, duration, and urgency) in the complex network in our partner hospital. Urgent patients must be scheduled for surgery with much shorter access delay targets than elective ones. Due to the uncertainty in surgical demand, we develop a distributionally robust optimization model that assumes only the moment information of surgical demand.

507