Informs Annual Meeting Phoenix 2018

INFORMS Phoenix – 2018

WD41

3 - Robust Countable-state Markov Decision Processes Saumya Sinha, University of Washington, Seattle, WA, 98195, United States, Archis Ghate Policy iteration is a standard method for solving robust Markov decision processes. For the countable-state case, however, convergence of the algorithm has not been established. Also, an as-is implementation of the method is not possible since it calls for an infinite amount of computation in each iteration. We present an approximate policy iteration algorithm that performs finitely implementable variants of policy evaluation and policy improvement. We prove that the value functions of the sequence of policies produced by this algorithm converge monotonically to the optimal value function. The policies themselves converge subsequentially to an optimal policy. 4 - Feedback-based Tree Search for Reinforcement Learning Emmanuel Ekwedike, Tencent AI Lab, Bellevue, WA, United States Emmanuel Ekwedike, Princeton University, Princeton, NJ, United States, Daniel Jiang, Han Liu Inspired by recent successes of Monte-Carlo tree search (MCTS) in a number of artificial intelligence (AI) application domains, we propose a model-based reinforcement learning technique that iteratively applies MCTS on batches of small, finite-horizon versions of the original infinite-horizon Markov decision process. We provide the first sample complexity bounds for a tree search-based reinforcement learning algorithm. 5 - Analyzing and Provably Improving Fixed Budget Ranking and Selection Algorithms Di Wu, Georgia Institute of Technology, 755 Ferst Drive, NW, Atlanta, GA, 30309, United States, Enlu Zhou We study the fixed budget ranking and selection problem, where the goal is to maximize the probability of correct selection (PCS) under a fixed simulation budget. First, we characterize the convergence rate of PCS for several classes of algorithms, and reveal that a constant initial samples size only amounts to a sub- exponential (or even polynomial) convergence rate. Then, we improve two state-of-the-art algorithms: OCBA from simulation optimization and successive rejects (SR) from best-arm identification. Our algorithms are guaranteed to achieve an exponential convergence rate, as is shown by finite-sample bounds on the PCS. Further, the improvement is validated using numerical experiments. n WD41 North Bldg 226C Practice - Decision Support Systems & Applications I Contributed Session Chair: Juri Yanase, Complete Decisions, LLC, 10517 Springbrook Ave., Baton Rouge, LA, 70810, United States 1 - A Minimum Cost Consensus Model for Social-network Group Decision-making Problems with Incomplete Linguistic Preference Relations Dong Cheng, Xi’an Jiaotong University, Xi’an, 710049, China, Zhili Zhou In this study, we propose a minimum cost consensus framework in social- network group decision-making with incomplete linguistic preference relations. First, a uninorm-based iterative procedure is presented to estimate the missing preference values. Next, we obtain the user weights by analyzing the tie strength and topology structure of their social network. Then, inconsistent users are identified through both the consistency and consensus measures. To help them reach a consensus with the minimum adjustment costs, an optimization-based consensus model is built to provide customized recommendations to them. Finally, the validity of the proposed method is verified by an example. 2 - Optimal Pricing and Ordering Strategy in a Single Supplier Group Purchasing Problem with a Newsvendor Framework Abdollah Mohammadi, UNC Charlotte, 532 Lex Dr., Charlotte, NC, 28262, United States, Ertunga Ozelkan Procurement cost account for a significant percentage of the total cost in a business entity. Group purchasing is a procurement strategy that helps companies save in their purchasing cost. In this research we are considering a group purchasing problem where a single supplier offers a quantity discount pricing schedule and retailers should decide about their price and ordering policy to maximize their profit in a single period problem. Through theoretical analysis we find out the optimal result condition and develop a heuristic algorithm to find out the near optimal results for a test problem. 3 - Blockchain Adoption under Uncertainty of Regulation Reza Alizadeh, The University of Oklahoma, 815 Russell Circle, E. Brooks St, Norman, OK, 73071, United States, Leili Soltanisehat Demand for decentralization of the organizational processes is getting strong throughout the world. The entity of the decentralization technology and the governmental regulations may differ among enterprises. We consider a decentralized IT system (Blockchain) inside an enterprise under uncertain governmental regulation and standards when there is a competitive tendency to adopt blockchain in the market.

4 - Shared Decision Making (SDM) as a Fast Emerging Field in the Interface of OR / MS and Medicine: Its Past, Present, and Likely Future Juri Yanase, Complete Decisions, LLC, 10517 Springbrook Ave., Baton Rouge, LA, 70810, United States, Evangelos Triantaphyllou, Zaina Qureshi The rapid evolution of medicine has resulted in multiple treatment options for many chronic conditions. These treatments are associated with different risk- benefit profiles, and preferences by individual patients. SDM provides a unique platform for patients and clinicians to collaboratively determine the best treatment option for individual patients. We review SDM’s history, the state-of- the-art and some challenges for the future. n WD42 North Bldg 227A Practice- Simulation and Optimization Contributed Session Chair: Julien Vaes, University of Oxford, Kings Cross, London, NW1 2DB, United Kingdom 1 - Data Driven Optimization and Statistical Modeling to Improve Meter Reading for Utility Companies Debdatta Sinha Roy, Robert H. Smith School of Business, University of Maryland, 7699 Mowatt Lane, 3330 Van Munching Hall, College Park, MD, 20742, United States, Christof Defryn, Bruce L. Golden, Edward Wasil Utility companies collect usage data from meters on a regular basis. Each meter has a signal transmitter that is automatically read by a receiver within a specified distance using radio-frequency identification (RFID) technology. In practice, there is uncertainty while reading meters from the planned routes of the vehicles. The RFID signals are discontinuous, and each meter differs with respect to the specified distance. These factors can lead to missed reads. We use data analytics, optimization, and Bayesian statistics to address the uncertainty. Simulation experiments using real data show that a hierarchical Bayesian model performs the best by designing improved routes for the vehicles. 2 - UNIPOPT: Univariate Projection-based Optimization without Derivatives Ishan Bajaj, Texas A&M University, College Station, TX, 77840, United States, Faruque Hasan We present a novel derivative-free framework UNIPOPT (UNIvariate Projection- based OPTimization) based on projecting all the samples onto a univariate space defined by summation of the decision variables. We illustrate that a univariate function (defined as lower envelope) exists on this space such that its minima is the same as that of the original function. The UNIPOPT framework identifies the points on the lower envelope and uses these samples to optimize it. UNIPOPT finds solutions within 1% of the global minima for 10-30% more problems compared to other solvers when applied on 485 constrained black-box problems. We also show the convergence of UNIPOPT to first order critical point. 3 - Design a Power Network for Charging Electronic Vehicles Ting Wu, Nanjing University, Department of Mathematics, No 22 Hankou Road, Nanjing, 210093, China, Cheng Zhu, Yasmina Maizi Electronic Vehicles contribute to a green environment in a smart city, they, however, raise a challenging problem for an existing power network to accommodate their charging facilities. This study aims to verify an upgraded power network via simulation models, providing managerial insights for adjusting existing power networks given a transportation network. 4 - Optimal Trade Execution Strategy under Volume and Price Uncertainty Julien Vaes, University of Oxford, Andrew Wiles Building, Radcliffe Observatory Quarter, Woodstock Road, Oxford, OX2 6GG, United Kingdom, Julien Vaes, The Alan Turing Institute, 96 Euston Road, Kings Cross, London, NW1 2DB, United Kingdom In the seminal paper on optimal execution of portfolio transactions, Almgren and Chriss define the optimal trading strategy to liquidate a fixed volume of a single security under price uncertainty. Yet sometimes, like in the power market, the volume to be traded can only be estimated and becomes more accurate when approaching a specified delivery time. We develop a model that accounts for volume uncertainty and show that a risk-averse trader has benefit in delaying trades. We demonstrate that the optimal strategy is a trade-off between early and late trades to balance risk associated to price and volume respectively.

508