Informs Annual Meeting Phoenix 2018

INFORMS Phoenix – 2018

SA40

3 - Testing Safety of Autonomous Vehicles Using Data-driven Rare-event Simulation Hongseok Namkoong, Stanford University, Stanford, CA, 94305, United States Despite recent progress in autonomous vehicles (AVs), rigorous tests are yet to be designed. Viewing this as a data-driven rare-event simulation problem, we first study a stylized i.i.d. random walk setting and give finite-sample minimax bounds and asymptotic rates for estimating the large deviation rate from data. Then, we implement a simulation framework (pseudo-reality) which can test the deep- learning based perception and control pipeline as a whole system. Use a learned behavior of environmental agents as our underlying model, we demonstrate our framework on a highway driving scenario, and show that rare-event simulation techniques can accelerate AV system evaluation. 4 - Ranking and Selection with High Dimensional Covariates Xiaocheng Li, Stanford University, Stanford, CA, United States, Zeyu Zheng Ranking and selection is concerned with making the best selection among many alternatives, whose unknown performances can be learned via sampling. Examples include selecting medicine and treatment regimes in healthcare systems and selecting online advertisements targeting Internet users. In this talk, we consider the settings in which the mean performance of each alternative depends on some observable high dimensional random covariates, where existing methods may become inefficient and even inapplicable. Our procedure effectively incorporate machine learning tools such as regularization and variable selection. Certain statistical guarantee is established under moderate assumptions. n SA39 North Bldg 226A Inventory Control Sponsored: Applied Probability Sponsored Session Chair: Eugene A. Feinberg, Stony Brook University, Stony Brook, NY, 11794-3600, United States Co-Chair: Mark S. Squillante, IBM Thomas J. Watson Research Center, IBM Thomas J. Watson Research Center, Yorktown Heights, NY, 10598, United States 1 - Asymptotic Optimality of Constant-order Policies in Joint Pricing and Inventory Control Models Linwei Xin, University of Chicago, 5807 S. Woodlawn Avenue, Chicago, IL, 60637, United States, Xin Chen, Alexander Stolyar We consider a joint pricing and inventory control problem with positive replenishment lead times. Although this problem has been extensively studied in the literature, the structure of the optimal policy remains poorly understood. In this work, we propose a class of so-called constant-order list-price policies. We prove that the best constant-order list-price policy is indeed asymptotically optimal as the lead time grows large, which is exactly the setting in which the problem becomes computationally intractable due to the curse of dimensionality. We also show that the best constant-order list-price policy can be computed effectively. 2 - Optimize Inventory Freshness in Supply Chain Zhengliang Xue, IBM T.J. Watson Research Center, Yorktown Heights, NY, 10598, United States, David D. Yao, Markus Ettl This paper studies the management of inventory freshness and optimization of the local inventory policy for all the members in a decentralized supply chain. We approximate the expected profit by the closed-form solutions, and use them to study the efficiency and security for all the transactions of selling and buying the fresh products at different ages of the product life-cycle. 3 - Equicontinuity Conditions for Markov Decision Processes with Application to Inventory Control Eugene A. Feinberg, Stony Brook University, Department of Applied Mathematics, and Statistics, Stony Brook, NY, 11794- 3600, United States, Pavlo O. Kasyanov, Yan Liang This talk discusses the equicontinuity condition and its generalization, the lower semi-equicontinuity condition, for Markov decision processes with average costs per unit time. These conditions imply the validity of average-cost optimality equations, convergence of discounted-cost relative value functions to average-cost value functions, and continuity properties of average-cost value functions. Periodic-review stochastic inventory models typically satisfy these conditions. This implies the validity of average-cost optimality equations for inventory models with average cost criteria and provides useful tools to derive structural properties of average-cost optimal policies.

4 - Stochastic Setup Cost Inventory Model with Backorders and Quasiconvex Cost Functions Yan Liang, Stony Brook University, A-149, Math Tower, Stony Brook, NY, 11794, United States, Eugene A. Feinberg This talk is concerned with a periodic-review setup-cost inventory model with backorders and holding/backlog costs satisfying quasiconvexity assumptions. We show that this model satisfies assumptions that imply the validity of optimality equations for discounted and average-cost criteria and the family of discounted relative value functions is equicontinuous. We establish two groups of results: (i) optimality of (s,S) policies for infinite-horizon problems under discounted and average-cost criteria, and (ii) convergence of optimal discounted lower thresholds and discounted relative value functions to their average-cost counterparts as the discount factor converges to 1. n SA40 North Bldg 226B Spatial Queueing and Matching Systems Sponsored: Applied Probability Sponsored Session Chair: Varun Gupta, The University of Chicago Booth School of Business, Chicago, Illinois Co-Chair: Ankur Mani, University of Minnesota, Minneapolis, MN 1 - Pooling Policies in Ridesharing Daniel Freund, Cornell University, 109 Lake St, Ithaca, NY, 14850, United States, Siddhartha Banerjee, Varun Gupta, Samitha Samaranayake The ascent of ride-sharing platforms like Lyft, Didi, and Uber triggered a line of research investigating queueing-theoretic models of such systems. The fundamental matching task of interest in this area is to match drivers and passengers. In this work, we take that analysis one step further to analyze shared- ride modes like Lyft Line. For these settings we need to not only match drivers and passengers, but also match passengers with each other. We propose several (asymptotically optimal) policies of interest and discuss their tradeoffs in non- asymptotic regimes. 2 - Matching Users Online: Balancing Matching Cost and Waiting Cost Sai Sandeep, CMU, Pittsburgh, PA, United States, Varun Gupta, Ravishankar Krishnaswamy In ride sharing platforms, the primary goal is to match drivers with riders to ensure that riders are paired to nearby drivers without waiting too long. This gives rise to an optimization problem where requests of two types arrive online, and the objective is to pair requests of different types so as to minimize sum of matching cost, and waiting cost. In this project, we consider this problem and its variant, also known as Min-Cost Perfect Matching with Delay (MPMD), survey the existing results and present new algorithms. 3 - Matrix Geometric Analysis of Polling Queues in Series Ravi Suman, University of Wisconsin-Madison, Madison, WI, 53726, United States, Ananth Krishnamurthy We analyze a network of queues where the customers are served at each station in a fixed cyclic exhaustive manner. Under Markovian assumptions for arrival and service times, we conduct an exact analysis using matrix geometric approach. Exact solutions for queue length distributions and mean waiting times are determined and the effect of buffer sizes on performance measures are explored. 4 - Walking vs. Waiting: Trade-offs when Pooling Resources Serving a Large Area Daniel F. Silva, Auburn University, 3301 Shelby Center, Auburn, AL, 36849, United States, Mathias A. Klapp We study the trade-offs between travel time and queue time in a two dimensional area, where stations are spatially separated. Customer demand appears randomly in a designated area, then each customer walks to a station to get service and then either walks to an exit or back to where they began. Examples include stadium concessions; department store check-outs; bars in a night club. We make assumptions about demand distribution, potential station locations and user preferences, then attempt to optimize a user utility function, subject to a budget on the available stations and servers. 5 - A Spacial Queueing Model for Evaluating Feasibility of Last-mile On-demand Public Transportation Service Alexander Vinel, Auburn University, 2479 Churchill Cir, Auburn, AL, 36832, United States, Daniel F. Silva We consider modeling approaches for analysis of last-mile on-demand public transportation service. Specifically, we are interested in enabling evaluation of whether specific demand profiles and other system parameters allow for efficient implementation of such a service. We develop a series of models employing spacially and temporally distributed queuing systems. We will present some analytical conclusions as well as simulation results.

15