INFORMS 2021 Program Book

INFORMS Anaheim 2021

SC35

2 - Data Augmentation for Object Detection Models: A Case Study Mohammad Noroozi, University of South Florida, Tampa, FL, United States, Ankit Shah Small object detection is a challenging problem found in many application domains. An imbalance with respect to any input property associated with the images can significantly affect the performance of the object detection models. In this talk, we will go over the data sampling and data augmentation techniques that can be used to better train deep learning models. We present our approach by augmenting a small object dataset by synthesizing images for object detection and demonstrate our results using the YOLOv4 deep learning model. SC35 CC Room 210A In Person: On demand Mobility: Operation and Competition General Session Chair: Daniel Vignon, University of Michigan, Ann Arbor, MI, 48109- 2125, United States 1 - Fleet Sizing for Automated Mobility-on-Demand Services Michael Hyland, University of California, Irvine, Irvine, CA, 92697-3600, United States Fleet sizing for automated mobility-on-demand (AMOD) services have significant cost and service quality implications for AMOD users as well as congestion and emissions implications for society. While this problem has received considerable attention recently, a powerful modeling approach has been overlooked in the literature and in practice—the time-dependent transshipment problem (TDTP). This study shows how the TDTP can be adapted to determine the optimal fleet size for an AMOD service considering both purchasing and operating costs. The study also compares the TDTP approach with the state-of-the-art minimum path cover formulation approach, in terms of their ability to provide a reasonable bound for minimum fleet sizing. Bounds from both approaches are then compared with the results of an agent-based stochastic dynamic AMOD simulation model. 2 - Should Ubers be Used as Flexible Shuttles? Partha S. Mishra, PhD Student, Northwestern University, Evanston, IL, United States, Sunil Chopra, Sebastien Martin, Karen Smilowitz Today, larger corporations are looking for fast and economic commute alternatives for their employees. While there has been a boom in the variety of options, which now range across shuttles, micro-transit services and transit-on-demand services, the right mix of modes of transit and their scheduling has become a challenging task. This paper evaluates how the choice of the trade-off between capacity, cost and speed affects the mode of transit and its schedule amidst inherent variability. In particular, we make the case for the use of transit-on-demand services in combination with regular shuttles. 3 - Competition and Congestion in the Ride-for-hire Market Daniel Vignon, University of Michigan, Ann Arbor, MI, United States Comparatively little attention has been paid to the welfare-effects of competition in the ride-for-hire market, especially with respect to congestion. Thus, we model and investigate competition in the ride-for-hire market, its effect on congestion and the resulting market equilibrium. We especially focus on competition between two ridesourcing platforms, on one hand, and between a ridesourcing and a taxi company on the other. The outcomes of both scenarios are compared and contrasted with each other and with the outcome from a monopolist ridesourcing platform. Then, we propose and evaluate welfare-enhancing regulations and discuss practical policy implications for the ride-for-hire market. SC36 CC Room 210B In Person: Innovative Transportation and Urban Planning General Session Chair: Yiling Zhang, University of Minnesota, Minneapolis, MN, 55455- 0141, United States 1 - Impact of Congestion on Mobility Choices: A Semiparametric Machine Learning Approach Aron Brenner, Student, MIT, Cambridge, MA, United States, Manxi Wu, Saurabh Amin In this paper, we propose a class of semi-parametric machine learning methods to predict travelers’ aggregate mode choices between public transit and driving using real-time data of congestion delay in a transportation network. Our prediction method combines a parametric binary mode choice model, and a class of non- parametric machine learning tools used for reducing the data dimensionality. Our

method also outputs a weight on each road segment, which reflects the impact of travel time delay of that segment on travelers’ mode choices. We apply this method to predict the ratio between driving and subway ridership in the San- Francisco Bay area, and achieve less than 2% RMSE during morning rush hours. 2 - Learning and Inference of Travelers’ Route Choice Preferences from Network Level Data Guarda Pablo, Carnegie Mellon University, Pittsburgh, PA, United States, Sean Z. Qian This study extends classical bilevel formulations to learn travelers’ multi-attribute utility functions from traffic data and conduct hypothesis testing on the parameters’ estimates. Via the formulation of a non-linear least squares problem and experiments on synthetic data, we showed that the parameters of the utility function were consistently recovered. The realization of a pseudo-convex objective function motivated the integration of normalized gradient descent with standard second order optimization methods used in prior literature. The theory was deployed at a large scale using real-world traffic data in Fresno, CA 3 - Analytics-Driven Inspection Operations for Post-Disaster Infrastructure Damage Assessment Mathieu Dahan, Assistant Professor, Georgia Institute of Technology, Atlanta, GA, 30309, United States, Andrew Lee, Saurabh Amin Delayed identification of storm-induced damages across coastal infrastructures often contributes to high economic and societal costs. To reduce the impact of damage, we address a critical knowledge gap by developing an analytics-driven approach for inspection of large-scale infrastructure networks in the aftermath of disaster events. Our approach integrates flexible diagnostic information from fixed and aerial sensors to design inspection crew routing strategies for rapid post- disaster damage identification. Using drainage network inspection data following 2017’s Hurricane Harvey, we show that prioritizing inspection based on damage indicators leads to significant cost and time savings. 4 - Incorporating Kinematic Wave Theory Into a Deep Learning Method for Traffic Speed Estimation Bilal Thonnam Thodi, PhD Student, New York University Abu Dhabi, Abi Dhabi, United Arab Emirates We present a kinematic waves-based Deep Convolutional Neural Network (Deep CNN) for estimating high-resolution traffic speed fields using sparse probe vehicle trajectories. Two key notions incorporate traffic physical constraints into the learning framework. Firstly, the use of anisotropic traffic kernels in the Deep CNN model an architecture modification aimed to explicitly capture the space-time correlations in free-flow and congested traffic. These correlations are guided by the Kinematic Wave Theory of traffic flow. Secondly, simulation-based training the use of simulated data as a surrogate to real-world data for training. This implicitly honors traffic physical constraints underlying the simulated data, and hence the simulation model. Speed field estimations for two real-world datasets show promising results. In Person: New Strategies and Technologies in the Operations/Finance Interface & Social Responsibility and Risk in Supply Chain General Session Chair: Sergio Camelo, Stanford University, Stanford, CA, 94305, United States Co-Chair: Parshan Pakiman, University of Illinois-Chicago, Chicago, IL, 60605, United States 1 - The Analysis of Blockchain-based Decentralized Exchanges Ruizhe Jia, Columbia University, NY, United States, Agostino Capponi We investigate the market microstructure of Automated Market Makers (AMMs), the most prominent type of blockchain-based decentralized exchanges. We show that the order execution mechanism yields token value loss for liquidity providers if token exchange rates are volatile. AMMs are adopted only if their token pairs are of high personal use for investors, or the token price movements of the pair are highly correlated. A pricing curve with higher curvature reduces the arbitrage problem but also investors’ surplus. Pooling multiple tokens exacerbates the arbitrage problem. We provide statistical support for our main model implications using transaction-level data of AMMs. 2 - Matching Platforms for Smallholder Supply Chains Sergio Camelo, Stanford University, Stanford, CA, 94305, United States, Dan Iancu, Joanne de Zegher We design a centralized platform that matches smallholder farmers with middlemen that provide transportation for their fruit. The platform sells the fruit and uses its revenue to pay both parties. Payments are designed to ensure that participating in the platform is more profitable for both farmers and middlemen than working outside of it. To model the profits that both parties could obtain SC38 CC Room 210D

26