Informs Annual Meeting 2017

MA59

INFORMS Houston – 2017

- Variance Reduction in Sequential Sampling for Stochastic Programming

3 - An Integrated Framework for Assisted Evacuation of Remote Habitats Jyotirmoy Dalal, Assistant Professor, Indian Institute of Management Lucknow, House No. 422, IIM Lucknow Campus, Lucknow, 226013, India, jyotirmoy.dalal@iiml.ac.in In the wake of a disaster, automobile based self-evacuation is not always an option for many remote, economically backward population. Even in case of an assisted-evacuation, a lack of proper road infrastructure makes the mass-transit movement unsafe during the disaster. We propose an integrated, centralized, time- and cost-effective mass evacuation and sheltering problem. Given the habitat and shelter locations, our model makes the necessary location, capacity and allocation decisions for a certain number of evacuee-pickup points under various supply and demand side constraints. 362F Joint Session RAS/Practice: Railroad Machine Learning & Component Failure Analysis Sponsored: Railway Applications Sponsored Session Chair: Roochi Mishra, BNSF Railway Company, Fort Worth, TX, 76131, United States, roochi.mishra@bnsf.com Co-Chair: Qing He, University at Buffalo, SUNY, Buffalo, NY, 14260, United States, qinghe@buffalo.edu 1 - Data-driven Railway Track Deterioration Model Qing He, University at Buffalo, SUNY, 313 Bell Hall, Buffalo, NY, 14260, United States, qinghe@buffalo.edu, Amjad Aref This study proposes an ensemble track deterioration model based on both stochastic mechanics modeling and Big Data machine learning techniques to capture the growth of defects, including both geometry defects and rail defects. The proposed model has been validated by real-world track data from North America Class I railroads. 2 - A Machine Learning Approach to Generate Track Surfacing Plan Based on Geometry Car Test Data Pooja Dewan, PhD, BNSF Railway, Fort Worth, TX, 76137, United States, pooja.dewan@bnsf.com In railroad, geometry car provides foot by foot track surface measurements as basic guide line for railway road master to make surfacing plan. Traditionally, we use rule based or threshold based method to indicate the area required to do maintenance. In this presentation, we will present a machine learning based approach to generate track surfacing plan from geometry car test data. Details from data preparation, training data collection to model implementation and deployment will be discussed. 3 - Bearing Failure Analysis Landon Smith, BNSF Railway, BNSF Railway, Fort Worth, TX, United States, landon.smith@bnsf.com, Hark Braren In transportation via rail, a roller bearing failure on a car in a train generally leads to catastrophic derailments. For many years, railroads have utilized Hot Box Detectors (HBD) to generate system alerts for bearings that approach a high temperature state. Although this method is somewhat effective, it does not catch ALL failing bearings, and these non-alerted bearings cause various issues impeding smooth operations, such as setouts, service interruptions, and derailments. In this presentation, we will discuss the technology used for timely identification of failing bearings as well as mapping the failure progression modes for different types of bearing failures. 4 - Prediction of Arrival Times of Freight Trafficon US Railroads using Support Vector Regression Daniel Work, University of Illinois, Chicago, IL, United States, dbwork@illinois.edu, William Barbour Variability of travel times on the United States freight rail network is very high, which poses operational challenges if the nature of variability is not predictable. This work proposes a data-driven approach to predict estimated times of arrival (ETAs) of individual freight trains based on the properties of the train, network, and conflicting traffic. The problem is posed as a series of location-specific machine learning regression problems and solved using support vector regression trained on over two years of data. We present details on feature engineering, findings on the most predictive factors for ETA, and preliminary ETA improvements exceeding 20% over baseline statistical methods. MA59

Guzin Bayraksan, The Ohio State University, 210 Baker Systems, 1971 Neil Avenue, Columbus, OH, 43210-1271, United States, bayraksan.1@osu.edu, Jangho Park, Rebecca Stockbridge We investigate the use of variance reduction techniques Latin Hypercube Sampling (LHS) and Antithetic Variates (AV) in sequential sampling for stochastic programming. Sequential sampling takes a sequence of solutions (obtained by any method) and assesses its quality by sequentially increased sample size. We update the theoretical results for LHS and perform computational experiments for both LHS and AV. Our results indicate that the use of AV and LHS in a sequential setting leads to tighter confidence intervals on the solutions obtained with reduced computational effort. 3 - Optimization-based Quantification of Input Uncertainty Huajie Qian, University of Michigan, Ann Arbor, MI, United States, hqian@umich.edu, Henry Lam We study the empirical likelihood approach to construct statistically accurate confidence bounds for stochastic simulation under nonparametric input uncertainty. The approach is based on positing distributionally robust optimization problems with suitably averaged divergence constraints to provide asymptotic coverage guarantees. We present the theory giving rise to the constraints and their calibration, and demonstrate how the approach compares to existing methods such as the Bootstrap and the Delta method in terms of computational effort and stability. 4 - Quantifying Uncertainty in Sample Average Approximation via the Empirical Likelihood Henry Lam, University of Michigan, 1205 Beal Avenue, Industrial & Operations Engineering, Ann Arbor, MI, 48109, United States, khlam@umich.edu, Enlu Zhou We discuss the use of empirical likelihood in the statistical uncertainty quantification of stochastic objectives and constraints, and its connection to robust optimization. We compare this approach to some classical methods in obtaining performance bounds for stochastic optimization and discuss other recent related works. 362E Reliant Populations and Disrupted Systems Sponsored: Transportation Science & Logistics Sponsored Session Chair: Elise Miller-Hooks, George Mason University, GMU, Silver Spring, MD, na, United States, miller@gmu.edu Co-Chair: Jyotirmoy Dalal, Indian Institute of Management Lucknow, Prabandh Nagar, IIM Road, Uttar Pradesh, Lucknow, 226013, India, jyotirmoy.dalal@gmail.com 1 - Real-time Transit Demand Estimation during Disruptive and Emergency Events Tuan Le, George Mason University, Fairfax, VA, United States, tle34@masonlive.gmu.edu, Vadim Sokolov, Kuilin Zhang, Hubert Ley, James Garner We present Bayesian models and methods for analyzing transit data in problems of inference about demand for transit system in real time. The goal is to develop short-term (a few hours) forecasts during disruptive events, such as flooding, tornadoes, blizzards and man-made emergencies. In this work, we solve a variant of the origin-destination flow problem. Our models allow for explicit representation of measurement uncertainties and incorporation of prior knowledge about demand patterns, e.g. historically observed origin-destination flows. Historical data is modeled as a prior, and noisy measurements of boardings and alighting are modeled via a likelihood. 2 - Modeling Transit-Communication Interdependencies for Resilience Computation Neza Vodopivec, George Mason University, Fairfax, VA, United States, nvodopiv@gmu.edu, Elise Miller-Hooks Resilience to transit disruptions depends not only on a technical system’s performance, but also on the ability of individuals to adapt to disruptions. To make well-informed decisions, individuals depend on effective communication. We introduce an integrated socio-technical system model capturing functional dependencies between a system’s engineered and social components. We present algorithms used to evaluate component criticality in system performance and identify mechanisms of failure and failure propagation paths. MA58

153