INFORMS 2021 Program Book

INFORMS Anaheim 2021

SC10

SC09 CC Room 303D In Person: Applied Probability and Machine Learning General Session Chair: Chang-Han Rhee, Northwestern University, Chicago, IL, 60613- 5364, United States 1 - A Proximal Bundle Type Method for Smooth and Nonsmooth Convex Optimization and Stochastic Programming Jiaming Liang, Georgia Institute of Technology, Atlanta, GA, 30308-1214, United States, Renato D. C. Monteiro This talk presents a proximal bundle (PB) method for solving convex smooth and nonsmooth composite optimization problems. Like other proximal bundle variants, PB solves a sequence of prox bundle subproblems whose objective functions are regularized composite cutting-plane models. Moreover, PB uses a novel condition to decide whether to perform a serious or null iteration which does not necessarily yield a function value decrease. Iteration-complexity bounds for PB are established for a large range of prox stepsizes. We further extend PB to the stochastic setting where the objective function only has stochastic first-order oracle. To the best of our knowledge, this is the first time that a proximal bundle variant has been shown to be effective to solve convex stochastic programming problems. 2 - Eliminating Sharp Local Minima From SGD with Truncated Heavy-Tailed Noise Chang-Han Rhee, Northwestern University Chicago, IL, 60613- 5364, United States The empirical success of deep learning is often attributed to SGD’s mysterious ability to avoid sharp local minima in the loss landscape, which are known to be associated with poor generalization. Recently, empirical evidence of heavy-tailed gradient noise was reported in many deep learning tasks; under the presence of such heavy-tailed gradient noise, SGD can escape sharp local minima, providing a partial explanation to the mystery. In this talk, we analyze a popular version of SGD where gradients are truncated above a fixed threshold. We show that this SGD is not only capable of escaping sharp local minima but also effectively eliminates sharp local minima from its training trajectory. We prove that, under appropriate structural conditions, the dynamics of this SGD with small learning rates closely resemble those of a Markov jump process that never visits any sharp minima. SC10 CC Room 304B In Person: Topics in Monte Carlo Methods and Rare Event Sampling General Session Chair: Michael Conroy, University of North Carolina, Chapel Hill, Carrboro, NC, 27510, United States 1 - Importance Sampling for Maxima on Trees We develop an unbiased and strongly efficient importance sampler for tail events of solutions to max-type stochastic fixed point equations that are constructed on weighted, marked Galton-Watson processes. These solutions are also describable as the all-time maximum of a branching random walk with a perturbation. The sampler is based on a representation of the tail events after a change of measure, generalizing non-branching representations that are standard in Cramer- Lundberg theory. Related to spine changes of measure often used in the branching process literature, the new measure tilts only one path of the walk, inducing a structure on the underlying branching process that suggests even more efficient algorithms to approximate tail events for branching random walks. 2 - Approximating Quasi-stationary Distributions with Interacting Reinforced Random Walks Adam Waterbury, UC-Santa Barbara, Santa Barbara, CA, United States We propose two numerical schemes for approximating quasi-stationary distributions (QSD) of finite state Markov chains with absorbing states. Both schemes are described in terms of certain interacting chains in which the interaction is given in terms of the total time occupation measure of all particles in the system. The schemes can be viewed as combining the key features of the two basic methods for approximating QSD originating from the works of Fleming and Viot (1979) and Aldous, Flannery, and Palacios (1998), respectively. I will describe the two schemes, discuss their convergence properties, and present some exploratory numerical results comparing them to other QSD approximation methods. Michael Conroy, University of Arizona, Tucson, AZ, 27510, United States, Mariana Olvera-Cravioto, Bojan Basrak, Zbigniew Palmowski

SC05 CC Ballroom E / Virtual Theater 5 Hybrid Strategies for Successfully Passing Tenure Track Sponsored: Minority Issues Forum Sponsored Session Chair: Zahra Azadi, University of Miami Herbert Business School, Coral Gables, FL, 33158, United States 1 - Strategies for Successfully Passing Tenure Track Zahra Azadi, University of Miami Herbert Business School, Coral Gables, FL, 33158, United States The purpose of this session is to bring visibility to assistant professors on the tenure track. Panelists, including the department chair, professor, and associate professor, will share their experiences. This panel discusses the tenure process and tips for a successful promotion. 2 - Panelist Wedad Jasmine Elmaghraby, University of Maryland, College Park, MD, 20742-1815, United States 3 - Panelist Eduardo Perez, Texas State University, San Marcos, TX, 78666, United States 4 - Panelist Iris V. Rivero, Rochester Institute of Technology, Rochester, NY, 14623-5603, United States SC06 CC Room 303A In Person: Humanitarian Operations and Disaster Management II General Session Chair: Christopher W. Zobel, Virginia Tech, Blacksburg, VA, 24061- 0235, United States 1 - Improving Fuel Terminal Throughput During Natural Disasters: A Discrete Event Simulation Approach Shraddha Rana, Massachusetts Institute of Technology, Cambridge, MA, United States, Jarrod D. Goentzel, Justin J. Boutilier During natural disasters the increased demand for fuel is met with distribution infrastructure and processing constraints. We use discrete event simulation to model the movement of gasoline from fuel terminals to retail stations, via tanker trucks, in Florida. Our objective is to identify bottlenecks in the downstream distribution network and quantify how various interventions can increase throughput of fuel to aid hurricane evacuation and relief activities. This tool is aimed be used by policy makers to build intuition on effectiveness of response strategies. We find that by improving process rates, the existing distribution infrastructure can handle high demand without requiring additional facility locations. 2 - Game Theoretic Algorithm for Decentralized Network Restoration Alireza Rangrazjeddi, University of Oklahoma, Norman, OK, 73019-1022, United States, Kash Barker, Andres David Gonzalez A sustainable society critically demands a reliable interdependent infrastructure network. Due to the existence of interdependency among various networks, systems are highly sensitive to any incapacitation. Although the traditional point of view focused on optimizing critical interdependent infrastructure networks considering centralized analysis, having one actor as a decision-maker in the system is broadly biased from the actual environment. Therefore, in this study, we address the concern of having multiple decision-makers in the system with various reward functions by proposing a decentralized game theory algorithm for network restoration in post-disaster situations.

17