INFORMS 2021 Program Book

INFORMS Anaheim 2021

ME42

3 - Approximations and Heuristics For Fast Security Constrained Optimal Power Flow Constance Crozier, Univ, of Colorado, Security constrained optimal power flow (SCOPF) extends the traditional optimal power flow problem to a two-stage stochastic optimization model. The problem seeks to find the lowest cost way to deliver power to consumers, while considering a set of pre-defined contingencies (such as a component outage). For large networks, with many contingencies, the resulting problem contains millions of variables. Given that the power flow constraints are non-convex, it may not be possible to reach a solution in the required timescale. In this talk, we will discuss the approximations and heuristics used by the CU Boulder team in the ARPA-E GO competition to reduce the complexity of the problem. ME42 CC Room 212B In Person: Deep Learning/Machine Learning I Contributed Session Chair: Meghna Maity, Kansas State University, Manhattan, KS, 66502, United States 1 - Optimal Policy Trees Jack Dunn, Interpretable AI, Cambridge, MA, United States, Maxime Amram, Ying Zhuo We propose an approach for learning optimal tree-based prescription policies directly from data, combining methods for counterfactual estimation from the causal inference literature with recent advances in training globally-optimal decision trees. The resulting method, Optimal Policy Trees, yields interpretable prescription policies, is highly scalable, and handles both discrete and continuous treatments. We conduct extensive experiments on both synthetic and real-world datasets and demonstrate that these trees offer best-in-class performance across a wide variety of problems. 2 - A Theoretical Framework for Data Science Brian Wright, Assistant Professor/Director Undergraduate Program, University of Virginia, Charlottesville, VA, United States As the first School of Data Science in the country, University of Virginia faculty in the School have created a framework for Data Science that will drive research activity and shape academic programs. This session will discuss this framework in detail. 3 - Can a Joint Model Assist Target Label Prediction? Conditions and Approaches Jaeyoung Park, University of Florida, Gainesville, FL, United States, Muxuan Liang, Xiang Zhong Multi-label datasets are common in many practical problems. In order to borrow information from auxiliary labels, classical approaches build a joint model to predict multiple labels simultaneously. However, a joint model may not necessarily lead to better prediction for the target label. In this work, we propose a framework to effectively utilize the hidden structure such as the hidden layers learned in the joint model to aid the prediction of the target label, even when the joint model is misspecified. Further, we propose a conditional-independence- targeted Neural Network (CITNN) aiming at efficient learning of predictive hidden structure. 4 - An Interpretable Deep Learning Model to Predict Symptomatic Knee Osteoarthritis Using Radiographs Maryam Zokaeinikoo, Cleveland Clinic, Cleveland, OH, United States, Xiaojuan Li, Mingrui Yang Early prediction of knee osteoarthritis (OA) may help initiate potential interventions or treatments sooner, which can result in preventing or delaying knee OA development. This study aims to develop a deep learning (DL) model to predict the incident symptomatic radiographic knee OA over the next 78 months using radiographs at baseline. The proposed model combines clinical variables, along with the attention and probability scores from our trained DL model to predict the knee osteoarthritis incident. The developed multi-modal model demonstrated an AUC of 0.76. Moreover, our model is interpretable, which can detect the potential abnormalities using the learned attention scores. 5 - A Holistic Approach to Mitigating Contamination in Food Supply Chain under Uncertainty Meghna Maity, Kansas State University, Manhattan, KS, United States Our work involves a holistic approach to make a perishable food supply chain resilient by identifying and optimizing the relevant factors that lead to bacterial contamination in food. Data-driven technologies such as machine learning are employed to filter the relevant parameters impacting food quality through various supply chains. Next, by integrating the Bayesian Network in Markov decision process models, we determine the optimal parameter settings to mitigate such contamination. We also use emerging blockchain technology to track the contamination and improve transparency throughout the supply chain.

ME44 CC Room 213B In Person: Pricing and Revenue Management Contributed Session Chair: Wei Sun, IBM T. J. Watson Research Center, Yorktown Heights, NY, 10598, United States 1 - The Impact of Order Fulfilment Services Provided by Online Marketplace Operators on Third-party Seller’s Performance Hao Su, University of Maryland-College Park, College Park, MD, United States, Martin E. Dresner This paper examines the impact of order fulfilment services provided by marketplace operators on third-party sellers’ competitiveness and product performance. We also examine factors that may moderate the relationship between using order fulfilment services provided by marketplace operators and a third-party seller’s competitiveness: the cost of services, the shipment size/weight, and the product value. 2 - The Impact of Uncertainty on a Broker’s Optimal Bidding Decisions in B2B Markets Ozden Engin Cakici, American University, Washington, DC, United States, Itir Z. Karaesmen We study a broker’s problem of matching a buyer with stochastic suppliers. The broker bids at each supplier. After the suppliers evaluate the bids, the broker learns the procurement quantity and then ships the items from each supplier to the buyer. When there is a single supplier the problem reduces to a new type of newsvendor problem. We study the impact of uncertainty on the optimal bids. We prove that the broker may or may not increase the bid when faced with uncertainty compared to a case with no uncertainty. We provide conditions under which it is optimal for the broker to bid at multiple suppliers. We numerically find that the broker’s expected profit decreases in the correlation between suppliers. 3 - Assortment Optimization for Online Multiplayer Video Games Fan You, PhD Candidate, University of Colorado-Boulder, Boulder, CO, United States, Thomas Vossen We consider the assortment optimization problem for a class of online multiplayer video games, where the in-game store has a unique structure with two sections, Featured and Just For You (JFY). All customers are offered the same Featured assortment whereas JFY is used for personalized recommendations. We model the choice of customers under the constrained mixture of nested logit model, and design a MILP formulation, as well as a FPTAS. We also propose a Lagrangian upper bound and a fast heuristic. We provide theoretical guarantees of the MILP formulation, the FPTAS as well as the heuristic algorithm. Numerical experiments show that our approaches perform well across a variety of settings. 4 - Ticket Pricing via Prescriptive Model Distillation Wei Sun, Researcher, IBM Research, Yorktown Heights, NY, United States, Max R. Biggs, Shivaram Subramanian, Youssef Drissi, Markus Ettl Powerful blackbox machine learning models often lead to complex policies which are difficult to verify and manage. Biggs et. al 2021 proposed a decision tree approach to extract revenue-maximizing pricing policies which are also interpretable by separating the counterfactual estimation and policy learning steps. We implement and test this method on airline ticket sales data. Results show that this method is capable of achieving significant improvement over current pricing policies with just a few rules. ME46 CC Room 213D In Person: Project Management Contributed Session Chair: Mohsen Mohammadi, University of Louisville, Chicopee, MA, 01020, United States 1 - Interrelationships Among Project Attributes and Time-phased Resource Patterns in a Project Portfolio Vishwanath Hegde, Professor, California State University-East Bay, Hayward, CA, United States, Zinovy Radovilsky We analyze the interrelationships among project attributes, durations, and time- phased resource allocation patterns in a portfolio of engineering projects. We estimate parametric models that capture unique duration/resource usage patterns from a longitudinal dataset spanning eleven years and analyze the link between the patterns and project attributes. Our research enhances the macro estimation of duration and resource requirements for incoming projects.

98