INFORMS 2021 Program Book

INFORMS Anaheim 2021

SD30

our analysis is to connect the evolution of a parameter to its limiting counterpart over Wasserstein space. Our analysis generalizes to soft Q-learning, which is further connected to policy gradient. 2 - Implicit Regularization of Sub-gradient Method in Robust Matrix Recovery: Don’t Be Afraid of Outliers Jianhao Ma, University of Michigan, Ann Arbor, MI, United States It is well-known that simple short-sighted algorithms, such as gradient descent, generalize well in the over-parameterized learning tasks, due to their implicit regularization. However, it is unknown whether the implicit regularization of these algorithms can be extended to robust learning tasks. In this work, we provide a positive answer to this question in the context of robust matrix recovery problem. We show that a simple sub-gradient method with a novel spectral initialization converges to the true low-rank solution efficiently, when it is applied to the over-parameterized l1-loss function without any explicit regularization or rank constraint. Moreover, by building upon a new notion of restricted isometry property, called sign-RIP, we prove the robustness of the sub- gradient method against outliers in the over-parameterized regime. SD29 CC Room 207C Undergraduate Operations Research Prize 1 Award Session Chair: Trilce Encarnacion, University of Missouri- St. Louis, Saint Louis, MO, 63121, United States SD30 CC Room 207D In Person: Multimodal Data Fusion for Healthcare Applications General Session Chair: Nathan B. Gaw, Georgia Institute of Technology, Scottsdale, AZ, 85258-2222, United States 1 - Statistical Inference for High-dimensional and Large-scale Data with Noisy Labels Hyebin Song, Pennsylvania State University, University Park, PA, United States In many classification applications, we are presented with data with partially observed or contaminated labels. One example of such an application is in the analysis of datasets from deep mutational scanning (DMS) experiments in proteomics, which typically do not contain non-functional sequences. In this talk, I will present statistical approaches and algorithms for analyzing noisy, high- dimensional binary data, demonstrating the optimality and scalability of our proposed methods. Finally, I will present an application of our methodology to inferring sequence-function relationships and designing highly stabilized enzymes based on large-scale DMS data. 2 - Cross Recurrence Analysis for Pattern Matching of Multidimensional Physiological Signals Adam Meyers, Doctoral Candidate, The Pennsylvania State University, University Park, PA, United States, Hui Yang, Mohammed Buqammaz Cross recurrence quantification analysis (CRQA), based on cross recurrence plot (CRP), is an effective method to characterize and quantify nonlinear interrelationships between pairs of time series. Despite its many advantages, CRQA has largely been unutilized for pattern mining of multidimensional, especially spatiotemporal, physiological signals. We present new methodology to visualize a patient-to-patient network where distance corresponds to pairwise patient dissimilarity based on CRQA statistics. This methodology is evaluated on real data consisting of 3D spatiotemporal vectorcardiogram signals from healthy and diseased patients. Experimental results show that certain diagonal line measures in the CRP, including our proposed measure characterizing maximum pairwise similarity between signals, are effective in distinguishing between patients.

SD27 CC Room 206B In Person: Network Optimization: Theory and Applications General Session Chair: Demetrios Vasilios Papazaharias, Buffalo, NY, 14214-2399, United States Co- Chair: Akhil Singla, Northwestern University, Evanston, IL, 60201, United States 1 - Conditional Value-at-Risk Shortest-Path Interdiction Di Nguyen, Clemson University, Clemson, SC, United States, Cole Smith We investigate an interdictor-evader shortest-path problem in which the interdictor attacks the network in advance and therefore only knows that the arc costs are uniformly distributed in given finite non-negative intervals. The impact of interdiction, i.e., the exact increase in an arc cost if interdicted, is known to both players. The evader, however, observes the arc costs in real time and chooses a shortest path when traversing the network. The interdictor seeks a solution that protects against the worst cases by maximizing the Conditional Value-at-Risk of the evader’s shortest-path cost. 2 - Resilient Network Flow Models Masoud Eshghali, University of Arizona, Tucson, AZ, United States, Pavlo Krokhmal In this talk we propose metrics of network resilience with respect to exogenous stochastic disruptions in the context of network flow models, such as the classical maximum network flow and minimum-cost network flow models. The proposed approach is based on stochastic programming and is inspired by concepts of modern risk theory. The properties of the resulting problems are discussed, and numerical studies are presented, including Benders decomposition based solution algorithms. 3 - A Fork-Join Decision Flow Network for Fact Checking in Social Media Akhil Singla, Northwestern University, Evanston, IL, United States, Seyed Iravani Social media works with a network of fact-checkers to identify misinformation on their platforms. They face the trade-off between “speed” and “accuracy” of detecting misinformation. We propose an POMDP to strike a balance between speed and accuracy in a fork-join fact checking network. 4 - Optimal Task Planning of Adversarial Games: An Integer Programming Approach Demetrios Vasilios Papazaharias, Buffalo, NY, 14214-2399, United States, Jose L. Walteros, Moises Sudit In this study we will focus on determining an optimal schedule of tasks in an adversarial setting. We model this problem within the guise of network interdiction, where the defender seeks to complete a schedule of tasks in the minimum amount of time. The attacker can expend some resources in order to delay the processing time of the defender’s tasks. The attacker seeks a minimum cost interdiction plan to ensure that the minimum completion time of the defender’s schedule is bounded by a given parameter. We first model this problem as a bilevel mixed integer program. We then propose a reformulation with respect to the extreme points of the defender’s polyhedron. Finally, we develop a decomposition algorithm to handle its exponential size. SD28 CC Room 207B In Person: Recent Advances in Data-Driven Nonconvex Optimization General Session Chair: Salar Fattahi, Berkeley, CA, 94702-2147, United States Chair: Jianhao Ma, University of Michigan 1 - Can Temporal-difference and Q-learning Learn Representation? A Mean-field Theory Yufeng Zhang, Northwestern University, Evanston, IL, United States Temporal-difference and Q-learning play a key role in deep reinforcement learning, where they are empowered by expressive nonlinear function approximators such as neural networks. At the core of their empirical successes is the learned feature representation. We aim to answer the following questions: When the function approximator is a neural network, how does the associated feature representation evolve? We prove that utilizing an overparameterized two-layer neural network, temporal-di erence and Q-learning globally minimize the mean-squared projected Bellman error at a sublinear rate. The associated feature representation converges to the optimal one. The key to

39