Informs Annual Meeting Phoenix 2018

INFORMS Phoenix – 2018

MA09

4 - Tri-level Optimization for Enhancing Interdependent Network Resilience Nafiseh Ghorbani Renani, The University of Oklahoma, 202 West Boyd St. room 124, Norman, OK, 73019, United States, Kash Barker, Andres David Gonzalez The reliability and resilience infrastructure networks are a growing concern among communities due to the occurrence of disruptions. Resilience is often thought of as the ability to withstand a disruption and recover quickly from the disruption. The vulnerability could describe the extend to which a network is disrupted, and recoverability could describe its trajectory of recovery. As such, we propose a tri-level protection/interdiction/restoration to represent decisions made (i) by a defender before a disruption to reduce network vulnerability, (ii) by an attacker to effectively disrupt the network, and (iii) by a defender after the disruption to enhance recoverability. n MA09 North Bldg 124B Large-scale Optimization Sponsored: Optimization/Nonlinear Programming Sponsored Session Chair: Aryan Mokhtari, Massachusetts Institute of Technology, Cambridge, MA, 02139, United States Co-Chair: Alejandro Ribeiro, Wynnewood, PA, 19096, United States 1 - Theory on the Absence of Spurious Optimality Cedric Josz, University of California, Berkeley, Berkeley, CA, United States, Ouyang Yi, Richard Zhang, Javad Lavaei, Somayeh Sojoudi We study the set of continuous functions that admit no spurious local optima which we term global functions. They satisfy various powerful properties for analyzing nonconvex and nonsmooth optimization problems. For instance, they satisfy a theorem akin to the fundamental uniform limit theorem in analysis regarding continuous functions. Global functions are also endowed with useful properties regarding composition of functions and change of variables. Using these new results, we show that a class of nonsmooth nonconvex optimization problems arising in tensor decomposition applications are global functions (e.g. with L1 objective functions). 2 - Generative Adversarial Networks (GANs) and Compressed Sensing Alexandros Dimakis, University of Texas at Austin, Austin, TX, United States The goal of compressed sensing is to estimate a vector from an underdetermined system of noisy linear measurements, by making use of prior knowledge on the structure of vectors in the relevant domain. For almost all results in this literature, the structure is represented by sparsity in a well-chosen basis. We show how to achieve guarantees similar to standard compressed sensing but without employing sparsity at all. Instead, we suppose that vectors lie near the range of a generative model, e.g. a GAN or a VAE. We show how the problems of image inpainting and super-resolution are special cases of our general framework. 3 - Fast Nonconvex SDP Solver for Large-scale Power System State Estimation (PSSE) Problem Hao Zhu, The University of Texas at Austin, 2501 Speedway, Austin, TX, 78712, United States Convex relaxation to a semi-definite program (SDP) has shown great success in power flow related problems of quadratic relations. High computational complexity of SDP solver however, challenges its large-scale application in real- time monitoring such as power system state estimation (PSSE). We will introduce an accelerated solver for large-scale PSSE by leveraging recent advances on non- convex SDP formulation that allows a lower-dimensional matrix search space. The accelerated gradient descent method is adopted to iteratively solve the resultant problem, at low per-iteration complexity thanks to the problem sparsity structure therein. 4 - Quantized Decentralized Consensus Optimization Ramtin Pedarsani, UC Santa Barbara, ECE Department, UCSB, Santa Barbara, CA, 93106, United States, Amirhossein Reisizadeh, Aryan Mokhtari, Hamed Hassani We consider the problem of decentralized consensus optimization, where the sum of n convex functions are minimized over n distributed nodes that form a connected network. We consider the case that the communicated local decision variables among nodes are quantized to alleviate the communication bottleneck in distributed optimization. We propose the Quantized Decentralized Gradient Descent (QDGD) algorithm, in which nodes update their decision variables by combining the quantized information received from the neighbors with their local information. We prove that under standard strong convexity and smoothness assumptions, QDGD achieves a vanishing mean solution error.

n MA10 North Bldg 125A Interface of Operations and Finance: Optimization and Game Theory Methods Sponsored: Manufacturing & Service Oper Mgmt Sponsored Session Chair: Zhen Liu, PhD, Daniel L. Goodwin College of Business, Benedictine University, 5700 College Road, Goodwin Hall Rm 366, Lisle, IL, 60532-2851, United States 1 - An Incentive Mechanism Based on Trade Credit in a Risk-averse Supply Chain under Asymmetric Information Lian Qi, Department of Supply Chain Management, Rutgers Business School, Newark, NJ, United States, Zhihong Wang, Zhen Liu we take the model under information symmetry as the benchmark; the trade credit incentive model under information asymmetry is constructed based on the principal-agent framework, and we obtain the optimal trade credit contract configuration and further deduce the optimal decision of the retailer. Then, we analyze the validity of the contract and the influence of the private information and risk-aversion coefficient on the contract parameters and the selling price. The study shows that when the degree of risk aversion is within a certain range, the reasonable trade credit contract designed by the supplier can effectively encourage the retailer to report its real sales cost. 2 - Depth-limited Solving in Imperfect-information Games Tuomas W. Sandholm, Angel Jordan Professor of Computer Science, Carnegie Mellon University, Pittsburgh, PA, 15213, United States, Noam Brown, Brandon Amos A key challenge in imperfect-information games is that states do not have defined values. So, depth-limited search algorithms used in single-agent and perfect- information settings do not apply. We introduce a principled way to conduct depth-limited solving in imperfect-information games by allowing the opponent to choose among multiple strategies for the remainder of the game at the depth limit. Each of the strategies results in a different set of values for leaf nodes. This forces an agent to be robust to different opponent strategies. On a 4-core CPU and 16 GB memory, we build a master-level heads-up no-limit Texas hold’em AI that defeats many prior top AIs. Previously this required a supercomputer. 3 - An Analytical Treatment of Dynamic Pricing and Learning with Menu Costs Zhen Liu, Assistant Professor, Benedictine University Goodwin College of Business, Goodwin Hall Room 366, 5700 College Road, Lisle, IL, 60532, United States Many firms face the problem of selling a given stock of products by a deadline. In their classic paper, Gallego and Ryzin (1994) show the optimal pricing policy is a function of the stock level and the length of the horizon. However, empirical studies show that the firms do not change their pricing policy as frequently as predicted, which leads to price stickiness. To better explain the discrepancy, we introduce menu costs into the problem and formulate it as a continuous-time impulse intensity control problem with Poisson-style demand processes. We obtain our optimal policy in threshold-type. Based upon this policy, we discuss approximation procedure for pricing policy under Bayesian learning. n MA11 North Bldg 125B Topics at the Interface of Finance, Operations and Risk Management Sponsored: Manufacturing & Service Oper Mgmt/iFORM Sponsored Session Chair: Dan Iancu Iancu, Stanford School of Business 1 - Risk Propagation in the Mortgage Supply Chain and the Financial Crisis Marco Yu Zhang, Illinois Institute of Technology, Chicago, IL, United States, John R. Birge Securitization has often been cited as the innovation in the mortgage supply chain that increased mortgage supply and the fragility of the network in leading to the financial crisis of 2008-2009. This supply-driven explanation, however, misses an important demand driver caused by drop in funding ratios increased risk-taking incentives for pension funds. This talk will describe a supply chain model to explain this phenomena and will present empirical support for the dominance of the demand effect.

122