Informs Annual Meeting 2017

TE08

INFORMS Houston – 2017

TE08

TE09

322B Disaster and Emergency Management Contributed Session Chair: Sandeep Grover, YMCA University of Science & Technology, Faridabad, India, groversandeep@hotmail.com 1 - Optimal Recovery Strategy for the Power Grid after a Cascading Failure Yan-Fu Li, Tsinghua University, N410, Shunde Building, Haidian District, Beijing, 100084, China, liyanfu@tsinghua.edu.cn, Yi-Fei Yuan Although heavily protected, electric power grids fail more frequent than expected. Severe failures, e.g. cascading failure, often result to significant damages to our society. Therefore it is of great importance to quickly repair the damaged power grid in the immediate aftermath of such failures. Whereas, the repair actions are cost-associated. In this study, we formulate an optimization problem for repair activity planning with the objective to minimize the total cost during recovery process, which includes the expense of repair actions and the economic loss due to load shedding. The proposed model is demonstrated on a Chinese provincial power grid. 2 - Network Planning Model for Food Aid Distribution During Chennai Floods: A Big Data Analytic Approach Ravi Shankar, Indian Institute of Technology Delhi, Department of Management Studies, Vishwakarma Bhavan, Shaheed Jeet Singh Marg, New Delhi, 110016, India, ravi1@dms.iitd.ac.in, Ashish Kumar Kaushal In this paper we design an effective last-mile food aid distribution network using big data analytics. We present location models to determine a set of distribution centres, where the food is directly distributed to the beneficiaries, during Chennai floods, 2015 in India. Our models take into account the consideration of all key stakeholders involved in the response system. 3 - Understanding Strategies Governing the Use of Social Data in a Disaster Response Operation Justin Taylor, University of Arkansas, 4207 Bell Engineering Center, 1 University of Arkansas, Fayetteville, AR, 72701, United States, jlt025@email.uark.edu, Ashlea Bennett Milburn The use of social data during disaster response has the ability to provide more information to emergency responders quickly, however the data is not guaranteed to be accurate. As such, responders are likely to have concern over the incorporation of social data and may limit the use of it by organizational policies or personal preferences. The research aims to help define what the current social data collection strategies are through interviews with emergency managers, as well as characterize the social data that are used during a disaster response operation. 4 - A Decision Support System for Location of Warehouses During the Preparedness Phase Ravi Shankar, Indian Institute of Technology Delhi, Delhi, India, ashishkiitd89@gmail.com, Ashish Kumar Kaushal This study integrates the fuzzy DEMATEL and goal programming model for solving the warehouse location problem during preparedness phase for effective disaster response. The managerial usefulness of this proposed method is that it allows decision makers to set multiple aspiration levels for the decision criteria. 5 - Early Forecasting of Flood Response Requirements for American Red Cross Mass Care Services Sanjeev Goyal, Assistant Professor, YMCA University of Science & Technology (Faridabad, India), NH-2, Sector 6, Mathura Road, In the United States, after a disaster resources are sent based on needs assessed at the local level. However, this local assessment can take time while local resources are preoccupied with immediate response activities. This delays the mobilization of personnel, resources, and supplies into the disaster area. We develop a machine learning method to estimate the required American Red Cross sheltering and feeding resources using data that would be readily available to disaster response managers on the first day of a flood. This enables more rapid mobilization and deployment and more effective provision of services to those affected by disaster. Faridabad, 121006, India, goyalsanjeev@pitt.edu, Kiatikun Louis Luangkesorn, Sandeep Grover

330A Machine Learning in Healthcare Sponsored: Health Applications Sponsored Session Chair: Yingfei Wang, Princeton University, Princeton, NJ, 08540-5233, United States, yingfei@princeton.edu 1 - The Behavior of Value of Information Policies in the Presence of a Locally Quadratic Belief Model Nana Kwabena Aboagye, Princeton University, Sherrerd Hall, Charlton Street, Apartment A, Princeton, NJ, 08544, United States, aboagye@princeton.edu, Warren B.Powell We study the problem of learning the unknown parameters of an expensive function where the true underlying surface can be well-approximated by a quadratic polynomial. We demonstrate that the well-known behavior of many learning policies for lookup table belief models no longer apply, and that the behavior when using a quadratic belief model exhibits a very special structure. This is the first paper to identify and exploit the behavior of learning policies in the presence of parametric belief models. We derive a simple rule that is much simpler to compute than a knowledge gradient policy that maximizes the one- step value of information and demonstrate its performance empirically. 2 - Machine Learning in Healthcare Nilakantan Sundara Raman Narasinganallur, K.J. Somaiya Institute of Management Studies & Research, B.602 Tulip Rachna Garden, Mulund Colony, Mumbai, 400082, India, nilakantan@somaiya.edu Healthcare is one of the new areas with potential for data analytics. With the generation of voluminous data, the use of big data analytics is essential for reaching useful conclusions in medical research. While past efforts in data analytics have been more in the area of statistical methods based on assumptions of linear data structures, machine learning provides an alternate basis to analyse for nonlinear relationships. The presentation will provide an overview of the use of machine learning in healthcare. 3 - Estimating the Cost-savings of Preventive Dental Services by Unsupervised Learning Ilbin Lee, Assistant Professor, University of Alberta, 3-23 Business Building, Edmonton, AB, T6G 2R6, Canada, ilee79@gatech.edu, Sean Monahan, Nicoleta Serban, Paul Griffin, Scott L. Tomar Dental caries is the most common chronic disease among U.S. children. Preventive services have been shown to be effective in preventing caries, but little is known about how the use of preventive services affects downstream care expenditures. We followed dental care utilization of nearly a million Medicaid children for 7 years and classified them using an unsupervised clustering algorithm. The results showed that expenditures were much lower for children who received preventive care than those who did not, and thus, preventive dental care can have substantial return on investments while also improving oral health outcomes. 4 - Adapting Kernel Mean Matching to Balance Observational Data in Health Care Martin Cousineau, PhD Candidate, McGill University, 1001 Sherbrooke St W, Montreal, QC, H3A 1G5, Canada, martin.cousineau@mail.mcgill.ca, Susan A. Murphy, Joelle Pineau, Vedat Verter, Gustavo Turecki We propose the use of kernel mean matching to compute a causal effect from observational data. We also propose an improved tuning procedure for kernel mean matching. 5 - Approximation Methods for Markov Decision Processes with Application to Clinical Trial Design Vishal Ahuja, Southern Methodist University, Cox School of Business, P.O. Box 750333, Dallas, TX, 75275, United States, vahuja@smu.edu, John R. Birge Multi-armed bandit problems, typically modeled as Markov decision processes (MDPs), exemplify the exploration vs. exploitation tradeoff. For many practical problems of interest, the state space is intractably large, rendering exact approaches to solving MDPs impractical. We propose a novel approximation approach that combines the strengths of multiple methods - grid-based state discretization, methods to improve approximation accuracy, and simulation, to obtain near-optimal policies for large scale MDPs with minimal added computational burden. Our numerical analysis shows our design to be almost as good as a fully optimal design and superior to existing heuristics.

412