Informs Annual Meeting 2017

SD17

INFORMS Houston – 2017

2 - Value of IoT in Spare Parts Logistics Networks Erhan Kutanoglu, University of Texas-Austin, OR/IE Graduate Program, Dept. of Mechanical Engineering, Austin, TX, 78712, United States, erhank@mail.utexas.edu, Ilyas Iyoob, Krishna T.Rekapalli We introduce a framework to evaluate the impact of continuous data feed on the maintenance condition of in-use equipment shared via the Internet of Things technology on a multi-echelon spare parts logistics network. We develop a simulation model and conduct extensive experiments to study the system-wide effects of different factors like shipment times, inspection intervals, replacement policies, etc. The simulation results help us analyze different policies as wells as the impact of IoT on several measures such as cost, inventory levels, and system efficiency under different settings. We introduce a metric that reflects the cost to benefit ratio for IoT investment. 3 - Collaborative Robots for Material Handling in Dynamic Environments Henry I. Ibekwe, GreenBerry Logistics, 10777 Westheimer, Suite 1100, Houston, TX, 77042, United States, hibekwe@startmail.com Modern state-of-the-art material handling systems are highly dynamic and complex requiring sophisticated automated facilities, advanced inventory management systems, and material moving equipment. There is constantly a need to optimize the process to reduce cost and increase output. This has led to interest in the application of adaptive, learning collaborative robots to perform repetitive and mundane tasks in contrast to the traditional fixed robot manipulator systems. We explore various models of collaborative robots that perform tasks alongside human personnel as a means to improve the operational efficiency and reduce the complexity of material handling in diverse environments. 4 - Channel Competition for Inventory with in an Omnichannel Ship- from-store Network: Examining Channel Sales Cannibalization Daniel Taylor, PhD Candidate, The Ohio State University, Fisher Implementing buy-online-ship-from-store (BOSS) tends to increase online sales. Without inventory adjustment, the policy can also result in a decline of in-store sales. Shifting inventory from an online distribution center to store-based fulfillment centers tends to increase in-store sales at those centers, but may place some online sales in jeopardy. This paper uses simulation and stochastic optimization to demonstrate these phenomena and to illustrate the effect of some BOSS inventory policies on channel service levels. College of Business, Columbus, OH, 43210, United States, taylor.465@osu.edu, Keely Croxton, A. Michael Knemeyer

Furthermore, we show the optimal reflection level can be derived as the fixed point that equates the long-run average cost to the holding cost. We also show the asymptotic optimality of this reflection control when it is applied to production-inventory systems driven by discrete counting processes. 4 - On the Control of Density-dependent Stochastic Population Processes with Time-varying Behavior Mark S. Squillante, IBM.Thomas J. Watson Research Center, Thomas J. Watson Research Center, 1101 Kitchawan Road, Yorktown Heights, NY, 10598, United States, mss@us.ibm.com, Yingdong Lu, Chai W. Wu We study a class of density-dependent stochastic population processes in which the rates that impact the density-dependent population structures and behaviors are time varying. Under a mean-field scaling, we extend classical results to show that such density-dependent stochastic population processes converge to a corresponding dynamical system and that the optimal control of such density- dependent stochastic population processes converges to the optimal control of the limiting dynamical system. We further investigate local (micro) and global (macro) behaviors within these processes and extend our results to support nearly completely decomposable stochastic process structures. 340B Stochastic OR Methods in Machine Learning Sponsored: Applied Probability Sponsored Session Chair: Jose Blanchet, Columbia University, New York, NY, 10027, United States, jose.blanchet@gmail.com 1 - Unbiased Monte Carlo Computations for Optimization and its Applications Yanan Pei, Columbia University, 506 W. 122nd St Apt 22, New York, NY, 10027, United States, yp2342@columbia.edu, Jose Blanchet, Peter W. Glynn We present general principles for the design and analysis of unbiased Monte Carlo estimators for optimal solutions of general stochastic optimization. Our estimators possess finite work-normalized variance under mild regularity conditions, such as uniqueness of optimizer and Lipschitz continuity of objective functions. With such unbiased estimators, it is convenient to implement estimations in parallel computations in practice. We apply our estimators to various settings of interest, such as regularized logistic regression and LASSO. Numerical experiments are also given to illustrate our result on real industry data. 2 - Robust Wasserstein Profile Inference: A Novel Approach Towards Choosing Regularization Parameters in Machine Learning Karthyek Murthy, Columbia University, New York, NY, United States, karthyek@gmail.com, Jose Blanchet, Yang Kang The objective of this talk is to introduce a novel, distributionally robust optimization based inference procedure called Robust Wasserstein Profile Inference (RWPI). The proposed procedure exploits connections between Empirical Likelihood, Distributionally Robust Optimization (DRO) and the theory of Optimal Transport. A key element of RWPI is the so-called Robust Wasserstein Profile function, a suitably scaled version of which, as we shall see, allows us to optimally choose tuning parameters for popular machine learning estimators such as generalized Lasso, regularized logistic regression, etc., without having to use the brute-force tuning approach of cross-validation. 3 - Doubly Robust Data-driven Distributionally Robust Optimization Fei He, Columbia University, IEOR.Department, Mudd 345, New York, NY, 10027, United States, fh2293@columbia.edu We formulate a double robust framework for optimal transport based distributionally robust optimization (DRO), and explore its applications in statistical learning. We apply robust optimization (RO) method to metric learning to learn a robust data-driven cost function and apply the cost function into DRO formulation, where we call it double robust DRO. DRO deals with finite sample uncertainty in model fitting; while RO metric learning takes care of the noisy estimation in constraints set. The double robust DRO provides a robust data- driven regularization for model fitting. Numerical experiments on real data sets shows that our method is superior comparing to its single robust counterpart. 4 - Data Driven Optimal Transport Cost Selection for Distributionally Robust Optimization Fan Zhang, Stanford University, Stanford, CA, United States, fzh@stanford.edu, Jose Blanchet, Yang Kang, Karthyek Murthy Recently, Blanchet, Kang and Murthy (2016) have shown that many machine learning algorithms, such as generalized Lasso, Support Vector Machines, and regularized logistic regressions, among many others, can be exactly represented as distributionally robust optimization (DRO) problems. The distributional uncertainty is defined as a neighborhood centered at the empirical distribution. We propose a methodology which learns such neighborhood in a natural data- driven way, thereby improving both theoretically and empirically (as we demonstrate) upon the classical algorithms. SD18

SD17

340A Optimal Control of Stochastic Systems Sponsored: Applied Probability Sponsored Session

Chair: Mark S Squillante, IBM Thomas J. Watson Research Center, Yorktown Heights, NY, 10598, United States, mss@us.ibm.com 1 - Electrical Vehicle Charging: A Queueing Approach Bert Zwart, Buddy Boldenlaan 44, Eindhoven, 5629RD, Netherlands, Bert.Zwart@cwi.nl We propose and analyse stochastic systems that model the performance of large- scale electrical vehicle charging systems. 2 - Revenue Management for Finite-buffered Tandem Queues Tonghoon Suk, IBM.Thomas J. Watson Research Center, 1101 Kitchawan Road, Yorktown Heights, NY, 10598, United States, tonghoon.suk@gmail.com, Xinchang Wang We study pricing policies for a tandem queueing system finite buffers. The problem is formulated as an infinite-horizon Markov decision process model maximizing the long-run average revenue. We propose a simple static policy and show that the policy is asymptotically optimal over all possible policies as the buffer size before the first station becomes sufficiently large. A variety of numerical studies are offered to test the performance of the simple static policy, and quite interestingly, we learn from the numerical results that the revenue obtained under the simple policy is quite close to the maximum obtainable

revenue even for moderate-sized buffer for the first station. 3 - On the Optimality of Reflection Control, with Production-Inventory Applications

Hengqing Ye, Assistant Professor, Hong Kong Polytechnic University, Kowloon, Hong Kong, lgtyehq@polyu.edu.hk, Jiankui Yang, David Yao

We study the control of a Brownian motion (BM) with a negative drift, so as to minimize a long-run average cost objective. We show the optimality of a class of reflection controls that prevent the BM from dropping below some negative level r, by cancelling out from time to time part of the negative drift; and this optimality is established for any holding cost function h(x) that is increasing in |x|.

106