INFORMS 2021 Program Book

INFORMS Anaheim 2021

ME16

Zhengtian Xu, The George Washington University, Washington, DC, 48105-2540, United States, Yafeng Yin, Xiuli Chao, Hongtu Zhu, Jieping Ye This study proposes a macroscopic fluid modeling framework to assist with strategic decision making of a platform for operating a large-scale on-demand app-based ride-hailing system. The framework captures the spatiotemporal characteristics of a ride-hailing system, and is flexible in representing control policies that a platform is implementing. It thus enables the analysis of a large- scale ride-hailing system with observable market responses and facilitates the optimization of the control polices. As a demonstration of the proposed framework, we customize it for a ride-hailing system operated by Didi Chuxing in a large city of China, and conduct an empirical study of the system. To our best knowledge, the proposed model is the first of its kind that offers a tractable way to support the analysis and optimization of large-scale ride-hailing systems. 2 - A Truthful Subsidy Scheme for Peer-to-peer Ridesharing Markets with Incomplete Information Amirmahdi Tafreshian, University of Michigan, Ann Arbor, MI, 48109-2125, United States, Neda Masoud Traffic congestion during peak periods has become a serious issue around the world, which is mainly due to the high number of single-occupancy commuter trips. Ridesharing platforms can present a suitable alternative for serving commuter trips. However, they face a major obstacle that prevents them from being a viable mode of transportation in practice; The users often provide tight time windows, which ultimately leads to a low matching rate. This study addresses this issue by introducing a subsidy scheme that allocates incentives to encourage a few carefully selected set of travelers to change their travel behavior. In order to implement this scheme for a ridesharing platform in the existence of private information, we propose an auction-based mechanism that guarantees truthfulness, individual rationality, budget-balance and computational efficiency. 3 - Network-level Impacts of Connected and Autonomous Vehicles on Traffic Congestion and Emission in a Mixed Traffic Environment Ehsan Kamjoo, Michigan State University, East Lansing, MI, United States, Fatemeh Fakhrmoosavi While many studies in the literature have investigated the impacts of connected vehicles (CV) and autonomous vehicles (AV) on traffic congestion and emission at the facility level, little is known about these impacts at large scales. Furthermore, there are uncertainties in the values of different parameters associated with these impacts, such as the extra vehicle-miles of traveled induced by AVs, technology cost of these vehicles, and possible reductions in the value of time of AV users, which have not been explored. Therefore, this study develops a stochastic framework and an algorithm to find the optimum market shares of CVs and AVs in a mixed traffic environment, consisting of human-driven vehicles without connectivity, CVs, and AVs, minimizing the system cost. Emission, travel time, and technology costs are considered as components of the system cost. 4 - Graph Based Approach to Real Time Metro Passenger Anomaly Detection Weiqi Zhang, Hong Kong University of Science and Technology, Hong Kong, Hong Kong Real-time anomaly detection of passenger flows in the metro system is very important to maintain the metro system’s normal operation and ensure passengers’ safety. We propose a novel abnormal passenger flow detection method based on smart card data. The method constructs a graphic model whose topological structure can capture the spatial distribution of anomalous passenger flow. It further incorporates external information (e.g. geographical information) to depict the latent passenger flow’s spatial dependence. A detection statistic is constructed by using graph community detection, which can be used for further signal selection and noise filter. It can be efficiently solved via a Min-Cut-based algorithm and can provide real-time solutions to anomaly detection and diagnosis. Preliminary experimental results demonstrate the efficiency of our method. 5 - Public Transportation Analysis via Tensor Decomposition and Spectral Clustering Nurettin Dorukhan Sergin, Arizona State University, Tempe, AZ, 85282-5544, United States, Hao Yan Automated fare collection systems record millions of transactions every day in major cities. These transactions, when analyzed, yield tremendous insight to public transportation decision-makers. We propose a model that undertakes two important analyses: discovering station similarities and detecting rare events in terms of hourly passenger inflow into stations. Our method combines tensor decomposition with spectral clustering and trains on a spatiotemporal tensor. We present our findings on smart card data collected over several months from Hong Kong’s subway system. 6 - A Macroscopic Model of Dockless Bike Sharing Systems Hongyu Zheng, Northwestern University, Evanston, IL, United States, Kenan Zhang, Yu (Marco) Nie This paper studies the design of a dockless bike sharing (DBS) system in a city using a parsimonious spatial model, in which the DBS service competes with walking and a generic motorized mode. Travelers choose one of the three modes according to their utility, which may depend on trip duration, access time and monetary cost. We show the access time to DBS is determined by the number of unique bike locations in the city, which is a function of the bike fleet size. The

model defines a supply-demand equilibrium that may be controlled by choosing the service price and the fleet size of the DBS system. We calibrate the model against empirical data collected in Chengdu, China, and test three counterfactual strategies: (i) profit maximization; (ii) market share maximization with non- negative profit; and (iii) social optimum. ME15 CC Room 201C In Person: Advances in Reinforcement Learning and Applications General Session Chair: Sadjad Anzabi Zadeh, University of Iowa, Coralville, IA, 52241- 2536, United States 1 - Optimal Dosing Protocol for Warfarin Using Deep Reinforcement Learning Sadjad Anzabi Zadeh, University of Iowa, Coralville, IA, 52241- 2536, United States Warfarin is a widely used anti-coagulant with narrow therapeutic range that makes it hard to optimally determine the daily dose. We show that deep Q- learning method can solve this problem and find the optimal dose for different patients with varying responses to the medication. Since the method requires a large volume of data, we employ a PK/PD model to simulate patients’ response to warfarin. 2 - Simple Agent, Complex Environment: Efficient Reinforcement Learning with Agent States Shi Dong, Stanford University, Stanford, CA, United States We design a simple reinforcement learning agent that, with a specification only of suitable internal state dynamics and a reward function, can operate with some degree of competence in any environment. We establish a regret bound demonstrating convergence to near-optimal per-period performance, where the time taken to achieve near-optimality is polynomial in the number of internal states and actions, as well as the reward averaging time of the best policy within the reference policy class, which is comprised of those that depend on history only through the agent’s internal state. Notably, there is no further dependence on the number of environment states or mixing times associated with other policies or statistics of history. ME16 CC Room 201D In Person: Participatory Specification of Trustworthy Machine Learning General Session Chair: Katia Guerra, University of North Texas, Flowermound, TX, 75028, United States 1 - Paving an Intentional Path Towards Inclusive Practices in AI Development Partnership on AI believes that working with communities affected by the deployment of AI/ML technologies is integral to their responsible development and mitigation of harm. To deepen our understanding of how inclusive public engagement approaches can help developers and researchers, PAI launched the Methods for Inclusion project. Multidisciplinary in nature, Methods for Inclusion draws from fields outside of computer science and technology, such as public planning & policy, education, public health, sociology, and community organizing which have grappled with questions of participation and inclusion for many decades. Drawing on semi-structured interviews with developers, data scientists, and researchers, as well as community advocates, the project explores the challenges present in conducting participatory design processes inclusively and equitably. 2 - Fair Performance Metric Elicitation Gaurush Hiranandani, Student, UIUC, Urbana, IL, United States, Narasimhan Harikrishna, Oluwasanmi Koyejo What is a fair performance metric? We consider the choice of fairness metrics through the lens of metric elicitation — a principled framework for selecting performance metrics that best reflect implicit preferences. The use of metric elicitation enables a practitioner to tune the performance and fairness metrics to the task, context, and population at hand. Specifically, we propose a novel strategy to elicit group-fair performance metrics for multiclass classification problems with multiple sensitive groups that also includes selecting the trade-off between predictive performance and fairness violation. The proposed elicitation strategy requires only relative preference feedback and is robust to both finite sample and feedback noise. Tina M. Park, Research Fellow (Methods for Inclusion), Partnership on AI, San Francisco, CA, United States

91