Informs Annual Meeting Phoenix 2018

INFORMS Phoenix – 2018

MA73

required the development of a number of new optimization models, characterizing their mathematical structure, and using this insight in designing algorithms to solve them. In our presentation, we focus on two particularly high- impact projects, an initiative to improve the allocation of docks to stations, and the creation of an incentive scheme to crowdsource rebalancing. Both of these projects have been fully implemented to improve the performance of Motivate’s systems across the country: Motivate has moved hundreds of docks in its systems nationwide and the Bike Angels program now aids rebalancing in San Francisco and NYC. In NYC, Bike Angels yields improvement comparable to that obtained through Motivate’s traditional rebalancing efforts, at far less financial and environmental cost. 2 - What’s Wrong with My Dishwasher: Advanced Analytics Improve the Diagnostic Process for Miele Technicians Segev Wasserkrug, IBM Haifa Research Lab, Haifa, 31905, Israel, Evgeny Shindin, Sergey Zeltyn, Martin Krueger, Yishai Feldman Miele, a leading appliance manufacturer, is looking to optimize the ways in which they solve customer problems quickly and efficiently. A crucial part of this task is precise diagnosis of faults, before and during technician visits. A correct diagnosis allows technicians to take with them the necessary parts and complete the repair with a minimal spending of time, effort, and spare parts. We created a decision- support system to help Miele optimize its service process, based on statistics learned from historical data about technician visits, containing both structured and unstructured (textual) data that had to be combined to create the probabilistic model. We used a novel process in which a semantic model informed the creation of the probabilistic model, as well as the analysis pipelines for the structured and unstructured data, combining expert knowledge with large heterogenous data. The results of a pilot study demonstrated a significant improvement of efficiency, concomitant with an increase of an already very high first-fix rate. n MA73 West Bldg 211B Crop Challenge Emerging Topic: INFORMS Special Sessions Emerging Topic Session Chair: Chris Tutino, Syngenta, 9 Davis Drive, Research Trinagle Park, NC, 27709, United States 1 - Using Data Analytics to Help Improve Agriculture in the 2019 Syngenta Crop Challenge in Analytics Greg Doonan, Syngenta, 2369 330 Street, Slater, IA, 50244, United States The 2019 Syngenta Crop Challenge in Analytics is a collaborative effort between Syngenta and the INFORMS Analytics Society that brings together experts in mathematics, agriculture and big data to demonstrate how data can be used to feed a growing global population with limited natural resources. Join Greg Doonan, head of novel algorithm advancement at Syngenta, to learn how you can use your data analytics skills to help scientists develop better crops. This year’s question asks participants to determine if corn hybrid performance data can be used to develop models that predict crop performance across a range of environmental scenarios. Understanding how a hybrid reacts to certain stresses can help breeders develop better seeds that allow farmers the potential to grow corn where they couldn’t before. The contest is unique because Syngenta gives participants real-world, proprietary data to create their models. The contest is open and submissions will be accepted until Jan. 18, 2019. Finalists will be announced in March and will be invited to present their entries at the 2019 INFORMS Conference on Business Analytics & Operations Research April 14-18, 2019 in Austin, Texas. The winners will be announced during the event. The first place winner will be awarded $5,000. Second place receives $2,500 and third place will receive $1,000.

n MA71 West Bldg 106C Joint Session ICS/ IOS Uncert: Adaptive/Sequential Sampling for Stochastic Programs Sponsored: Computing Sponsored Session Chair: Yongjia Song, Clemson University, SC, 29634, United States 1 - Variance Reduction in Sequential Sampling for Stochastic Programming Jangho Park, PhD Candidate, The Ohio State University, 1971 Neil Avenue, Columbus, OH, 43210-1271, United States, Rebecca Stockbridge, Guzin Bayraksan We investigate variance reduction techniques Latin Hypercube Sampling (LHS) and Antithetic Variates (AV) in sequential sampling for stochastic programming. Sequential sampling takes a sequence of solutions and assesses their quality by sequentially increased samples. We show conditions under which the procedures stop in a finite number of iterations and are valid. We computationally compare LHS and AV in sequential and non-sequential settings. Our results indicate that while both are effective, LHS dominates in the non-sequential setting and AV gains an advantage in the sequential setting. 2 - Adaptive Sequential Sampling for Stochastic Programs Yongjia Song, Clemson University, 211 Fernow Street, 264 Freeman Hall, Clemson, SC, 29634, United States, Raghu Pasupathy In this talk, we propose a new sequential sampling approach that unifies existing literature on the sequential sampling approach and the retrospective approximation approach for solving stochastic programs. Specifically, we stipulate schedules of sample size used for solving each sample-path problem at each iteration as has been done in the sequential sampling literature. In addition, we choose optimality gap tolerance adaptively, making it a factor of the sample error corresponding to the current iterate in the sample-path problem. We show that these choices are optimal under a precisely defined criterion. Numerical results on two-stage stochastic linear programs will be reported. 3 - Adaptive Sampling with Norm and Inner Product Tests for Smooth Stochastic Optimization Raghu Pasupathy We characterize the sample-path behavior of stochastic gradient descent (a.k.a. stochastic approximation). Such characterization leads naturally to adaptive sampling structures within stochastic gradient descent. We will discuss two specific variations of adaptive sampling SGD that use what has recently been called the norm test and inner product test, along with sample-path complexity results. Numerical results routinely demonstrate better practical performance while achieving sample-path convergence rates that are arbitrarily close to the best possible. 4 - Stochastic Algorithms for Conditional Stochastic Optimization Shuotao Diao, University of Southern California, Los Angeles, CA, 90007, United States, Suvrajeet Sen Conditional stochastic optimization models arise when the uncertainties involve correlated latent random effects. This presentation discusses several algorithms based on first-order methods in the context of nonparametric regression (e.g. k- NN). We study the asymptotic convergence of these algorithms. Joint Session Wagner/Practice: Daniel H. Wagner Prize for Excellence in Operations Research Practice Emerging Topic: Daniel H. Wagner Competition Emerging Topic Session Chair: Patricia Neri, SAS Institute, Inc., 104 Grandtree Ct., Cary, NC, 27519, United States 1 - Analytics and Bikes: Riding Tandem with Motivate to Improve Mobility Daniel Freund, Cornell University, 109 Lake St, Ithaca, NY, 14850, United States, Shane Henderson, David B. Shmoys, Eoin O’Mahony Bike-sharing systems are now ubiquitous across the U.S. We have worked with Motivate, the operator of the largest such systems, including in New York, Chicago and San Francisco, to innovate data-driven approaches for bike-sharing. With them we have developed methods to improve their day-to-day operations and also provide insight on central issues in the design of their systems. This work n MA72 West Bldg 211A

145