INFORMS 2021 Program Book

INFORMS Anaheim 2021

MB28

2 - Ecole: A Library for Learning Inside MILP Solvers Antoine Prouvost, Polytechnique Montréal, Montreal, QC, H3C 3A7, Canada, Justin Dumouchelle, Maxime Gasse, Didier Chételat, Andrea Lodi We describe Ecole (Extensible Combinatorial Optimization Learning Environments), a library to facilitate integration of machine learning in combinatorial optimization solvers. It exposes sequential decision making that must be performed in the process of solving as Markov decision processes. This means that, rather than trying to predict solutions to combinatorial optimization problems directly, Ecole allows machine learning to work in cooperation with a state-of-the-art a mixed-integer linear programming solver that acts as a controllable algorithm. Ecole provides a collection of computationally efficient, ready to use learning environments, which are also easy to extend to define novel training tasks. 3 - Mixed Integer Linear Optimization Formulations for Learning Optimal Binary Classification Trees Brandon Alston, Rice University, Houston, TX, United States, Hamidreza Validi, Illya V. Hicks Decision trees are powerful tools for classification and regression that attract many researchers working in the burgeoning area of machine learning. A binary classification tree is a special type of classification trees in which each branching node has exactly two children. An optimal binary classification tree can be obtained by solving a biobjective optimization problem that seeks minimizing (i) number of misclassified datapoints and (ii) number of branching nodes. In this paper, we propose three MILO formulations for designing optimal binary classification trees: two flow-based formulations and a cut-based formulation. In this talk, we propose three mixed integer linear optimization (MILO) formulations for designing optimal binary classification trees: two flow-based formulations and a cut-based formulation. 4 - Influence-Coverage Optimization Problem Majid Akhgar Farsani, Oklahoma State University, Stillwater, OK, United States, Juan Sebastian Borrero We introduce the problem of influence-coverage optimization, where the Influence Optimization and the Maximum Coverage problems are merged into one network optimization problem. In this problem, we have a social network where nodes get activated either by their active in-neighbors or by getting at least a point of their piecewise linear paths covered by external factors. Depending on the application, external factors could be facilities, flyers, billboards, Geo-fences, street art paintings, campaigns, social movements and so on. The aim is to minimize the time until all nodes are active, which depends on the locations of external factors and the value of in-neighbors influence rates. In this regard, we apply an exact approach as well as heuristic approaches providing both decent lower and upper bounds for the proposed MIP problem. In Person: Office of Naval Research Sponsored Talks/Recent Advances in High-Order Methods General Session Chair: David Phillips, U.S. Naval Academy, Annapolis, MD, 21401, United States Co-Chair: Soomin Lee, Yahoo! Research, Sunnyvale, CA, 94087, United States 1 - A Stochastic Newton Algorithm for Distributed Convex Optimization Brian Bullins, Toyota Technological Institute-Chicago, Chicago, IL, United States, Kumar K. Patel, Ohad Shamir, Nathan Srebro, Blake Woodworth We propose and analyze a stochastic Newton algorithm for distributed convex optimization. At the heart of our approach is recent work showing that quadratic objectives can be optimized to high accuracy using a parallel algorithm with only a single round of communication. Our algorithm expresses the Newton update as the solution to a quadratic problem which we optimize using stochastic gradients and stochastic Hessian-vector products for the objective, both of which can typically be computed efficiently. We analyze our method for quasi-self- concordant objectives (e.g., logistic regression), and demonstrate that it can in some instances achieve faster convergence rates than comparable first-order methods while requiring less communication and a similar amount of computation. MB28 CC Room 207B

2 - The Optimization Portfolio Supported by the Office of Naval Research: Current and Future Initiatives David Phillips, U.S. Naval Academy, Math Department, Annapolis, MD, 21401, United States In this talk, we will discuss different research directions in optimization and analytics the Office of Naval Research supports and encourages. We discuss examples of ongoing projects that illustrate the breadth of the basic research we support as well as the types of applied projects the Navy is interested in. MB29 CC Room 207C In Person: Transportation Science and Technology General Session Chair: Nilay Noyan, Amazon, Seattle, WA, 98109-5314, United States Co-Chair: Mauricio C. Resende, Amazon.com, Inc., Seattle, WA, 98109- 5314, United States 1 - Mitigating Spot Market Premiums Idil Arsik, Amazon, Seattle, WA, 98109, United States, Philip Kaminsky, Tara Mardan Amazon uses both committed capacity acquired weeks ahead of use, and a spot market to procure truckload capacity. The committed capacity is secured in the form of blocks of driver shifts with pre-specified start times and starting/ending location, and several hours to a day ahead of time, specific routes are assigned to these shifts. The delay in this assignment allows Amazon to effectively adapt to changing demand, but it also delays sending loads that are not planned into routes to the spot market, leading to higher premiums. We present a novel approach that is designed to mitigate this risk by offering the loads in the spot market at a reduced price while simultaneously considering them for routing. 2 - Block Time Planning with Atmospheric Wind Information Xiaofeng Wei, Amazon / Air Science and Technology, Bellevue, WA, United States, Rui Sun, Na An In today’s Cargo airline operations, block times are playing a critical role in the entire network, from planning fleet schedules, fuel cost and CO2 emissions to optimizing package shipment flows. Block time is composed of three components: taxi out time, flight time and taxi in time. In this paper, we develop forecast models using machine learning tools to predict each component of block time separately. Our model can apply the flight and taxi time patterns we learn from the existing OD pairs to forecast the flight and taxi times for new OD pairs and takes into account seasonal atmospheric wind patter for flight time. We test our forecast model using historical on-time performance data, and show that our model results outperform the benchmark forecast results compare to the other block time forecast methods. MB30 CC Room 207D In Person: Closing the Analytics Talent Gap General Session Chair: Jennifer Priestley, Kennesaw State University, Kennesaw, GA United States 1 - Closing the Analytical Talent Gap: What Analytics Professionals Should Know About Working With Universities (But Don’t) Jennifer Priestley, Kennesaw State University, Kennesaw, GA, 30144, United States, Robert Joseph McGrath Managers of Analytical Teams deal every day with the shortage of people with deep computational skills who are also capable of communicating results to non- technical audiences. As professors, we are asked almost daily from practitioners in the private sector - We have a project how do we reach out to your students? If we do research together who owns it? How much does all of this cost? At the same time, we hear from our academic colleagues - How can I bring a “real” analytical project in the classroom? Is what I am teaching aligned with the demands of the market? How do I start a conversation with a company? This session seeks to answer these questions and provide a tangible set of tools to establish industry-university collaborations. Case studies on working with universities from Equifax, The Home Depot, The Southern Company, and Shaw Industries will be integrated into the session.

52