2016 INFORMS Annual Meeting Program

TB86

INFORMS Nashville – 2016

5 - Optimizing Procurement Of High-value Medical Products In A Health-care Network Parimal Kulkarni, Manager, Supply Chain Analytics, BJC Healthcare, 8300 Eager Rd, Suite 500 D Mailstop 92-92-277, St Louis, MO, 63144, United States, pskf44@umsl.edu Parimal Kulkarni, Manager, Supply Chain Analytics, University of Missouri, St.Louis, One University Blvd, St Louis, MO, 63121, United States, pskf44@umsl.edu, L. Douglas Smith, Glen Moser We use MILP optimization and simulation in concert to develop procurement strategies for high-value medical supplies in a health-care network. With a multi- objective MILP model, we determine product groups to be purchased from alternative vendors to achieve quantity discounts while maintaining diversity of supply. Considered are physician preference, budgetary limits, and scorecards of vendor performance on several dimensions. Discrete-event simulation is used iteratively to test procurement solutions and help set the MILP constraints to keep risk at acceptable levels.

5 - Pricing Tax Return For Students That Opt-out From Using School Bus Hernan Andres Caceres Venegas, PhD Student, University at Buffalo - SUNY, 342 Bell Hall, University at Buffalo, Buffalo, NY, 14260, United States, hernanan@buffalo.edu, Rajan Batta, Qing He School districts are often mandated to provide transportation but can encounter ridership that varies between 22-72 percent. Consequently, buses run with unused capacity over long routes. We explore the scenario where students are compensated for giving up the option to ride a bus, in an effort to reduce the overall cost of the system. Mathematical formulations for this problem are developed and analyzed. Results from a case study along with algorithmic computational results will be presented. TB86 GIbson Board Room-Omni Monte Carlo Methods for Multi-stage Decision Making under Uncertainty Sponsored: Artificial Intelligence Sponsored Session Chair: Michael Fu, University of Maryland, mfu@isr.umd.edu 1 - Back To The Future: Google Deep Mind, Alpha Go & Monte Carlo Tree Search Michael Fu, University of Maryland, College Park, MD, 20742, United States, mfu@rhsmith.umd.edu In March 2016 in Seoul, Korea, Google DeepMind’s AlphaGo, a computer Go- playing program, defeated the reigning human world champion Go player, a feat far more impressive than previous computer programs victories in chess (Deep Blue) and Jeopardy (Watson). Due to the sheer combinatorial nature of the number of possibly game configurations, at the heart of all computer Go-playing algorithms is Monte Carlo tree search based on an upper confidence bound (UCB) algorithm that traces its roots back to an adaptive multi-stage sampling algorithm for estimating the value function in finite-horizon Markov decision processes (MDPs). We describe this algorithm and the main ideas behind AlphaGo. 2 - Cumulative Prospect Theory Meets Reinforcement Learning: New Monte Carlo Algorithms Cheng Jie, University of Maryland, cjie@math.umd.edu, Prashanth L.A., Michael Fu, Marcus Steve, Csaba Szepesvari We bring cumulative prospect theory (CPT) to a risk-sensitive reinforcement learning (RL) setting and present Monte Carlo simulation-based algorithms for both estimation and optimization. The estimation scheme uses the empirical distribution to estimate the CPT-value of a random variable. The optimization procedure is based on simultaneous perturbation stochastic approximation (SPSA). Both theoretical convergence results and numerical experiments are provided. 3 - Weighted Bandits Or: How Bandits Learn Distorted Values That Are Not Expected L. A. Prashanth, University of Maryland, College Park, MD, 20742, United States, prashla@umd.edu, Aditya Gopalan, Michael Fu, Steve Marcus We formulate a novel multi-armed bandit setup, where the arms’ reward distributions are distorted by a weight function. The distortions are motivated by models of human decision making that have been proposed to explain commonly observed deviations from conventional expected value preferences We study two representative problems in this setup: The classic K-armed bandit setting and the linearly parameterized bandit setting. In both settings, we propose algorithms that are inspired by UCB, incorporate reward distortions and exhibit sub-linear regret assuming Holder-continuous weights. We provide empirical demonstrations of the advantage due to using distortion-aware learning algorithms.

TB79 Legends G- Omni

Opt, Stochastic II Contributed Session

Chair: Hernan Andres Caceres Venegas, Ph.D. Student, University at Buffalo - SUNY, 342 Bell Hall, University at Buffalo, Buffalo, NY, 14260, United States, hernanan@buffalo.edu 1 - Efficient Solving Of Multi-stage Mixed-integer Stochastic Problems Under Mean-dispersion Distributional Information Krzysztof Postek, PhD Candidate, Tilburg University, Warandelaan 2, Tilburg, 5037 AB, Netherlands, k.postek@tilburguniversity.edu, Ward Romeijnders, Dick den Hertog, Maarten H van der Vlerk We propose a solution method for multi-stage robust optimization and stochastic programming problems under distributional uncertainty, when the means and mean absolute deviations of the parameters are known. Using new theoretical results we show for problems with integer recourse how to construct good convex approximations with known performance bounds and how to solve these problems efficiently. Our approach gives insights into the performance of the various recourse rules, the value of distributional information, and the trade-offs between different variants of the objective function (worst-case, worst-case expected, best-case). 2 - Multi-project Scheduling With Multi-mode Resource Constrained Under Uncertainty Berna Dengiz, Professor, Baskent University, Eskisehir Road 20th Km, Baglica Campus, Ankara, 06530, Turkey, bdengiz@baskent.edu.tr, Serdar Soysal In this study, we address a resource constrained project scheduling problem including uncertainties in resource usage rate in a multi-project environment. The activities of each project have alternative resource usage modes. Resources are dedicated to all projects considering their dedication policy. The projects involve finish to start zero time lag, nonpreemptive activities and limited renewable and nonrenewable resources. In this study, the optimal dedication of resource capacities to the projects and minimum value of weighted tardiness over all projects will be determined by proposed solution approach. 3 - Stochastic Integer Programming With Endogenous Uncertainty In Open Access Outpatient Clinic Appointments Scheduling Amarnath Banerjee, Associate Professor, Texas A&M University, 4041 Engineering Technology Building, 3131 Tamu, College Station, TX, 77843-3131, United States, banerjee@tamu.edu, Yu Fu This study develops a two-stage Stochastic Integer Programming (SIP) model to solve the online outpatient scheduling problem. The model considers different types of patients and uncertain factors in system throughput, no-show, cancellation and lateness. A modified L-shaped algorithm is designed to handle the endogenous uncertainty brought by these factors and solve the SIP model. The analysis method and solution algorithm can be applied to two-stage SIP Yiming Yao, Lawrence Livermore National Laboratory, 7000 East Avenue, L-181, Livermore, CA, 94550-9234, United States, yao3@llnl.gov, Vic Castillo, Andrew Mastin, Carol A Meyers, Deepak Rajan We present a two-stage stochastic mixed integer programming model that minimizes enterprise risk subject to supply, demand, capacity and other constraints, with the consideration of uncertainty in some parameter values. We describe risk measurement and uncertainty characterization in the application context. Finally, we describe the model’s implementation in the open source optimization modeling language PYOMO/PYSP. model with simple recourse function satisfying certain properties. 4 - A Stochastic Mixed Integer Programming Model For Risk Minimization

291