Informs Annual Meeting Phoenix 2018

INFORMS Phoenix – 2018

MC10

Donald Goldfarb We present several practical Gauss-Newton (Levenberg-Marquardt) methods for solving nonconvex optimization problems that arise in training deep neural networks involving enormous numbers of variables. Numerical results are presented to demonstrate the effectiveness of our proposed methods. 2 - Atomic SGD: Communication-efficient Learning via Atomic Sparsification Zachary Charles, UW-Madison, 1415 Engineering Drive, Madison, WI, 53703, United States, Dimitris Papailiopoulos Distributed implementations of mini-batch SGD are often beset by communication bottlenecks attributed to the large gradients that are communicated between compute nodes. One possible remedy is to sparsify these gradients. We show that such a sparsification can be applied to any atomic decomposition of the gradient, e.g., element-wise, spectral-wise, etc. Given a sparsity budget, we find a closed-form expression for the sparsification scheme that minimizes variance, which controls convergence. We show that methods like QSGD and TernGrad are special cases of our method, and argue that spectral sparsification of gradients can lead to significantly less communication costs compared to QSGD. 3 - Analysis of Scaled Memoryless BFGS on a Class of Nonsmooth Convex Functions Azam Asl, 251 Mercer St., New York, NY, 10012, United States, Michael Overton L-BFGS can be used with or without ``scaling”; the use of scaling is normally recommended. In this paper we analyze L-BFGS with one update with scaling on the nonsmooth function, $f(x) = a\vert x^{(1)}\vert + \sum_{i=2}^{n} x^{(i)}, \quad \mathrm{where} ~~x \in \R^{n}$. We show that if $a\ge \sqrt{4(n-1)}$, in the limit the the absolute value of the normalized search direction generated by this method converges to a constant vector, and if $a$ is larger than a quantity dependent only on the Armijo parameter, then the iterates converge to a non- optimal point $\bar x$ with $\bar x^{(1)}=0$, although $f$ is unbounded below. 4 - A Stochastic Trust Region Method Rui Shi, Lehigh University, Bethlehem, PA, United States In this talk, we present a new stochastic trust region method, deemed TRish, for solving stochastic and finite-sum minimization problems. We motivate our approach by illustrating how it can be derived from a trust region methodology. However, we also illustrate how a direct adaptation of a trust region methodology might fail to lead to general convergence guarantees. Hence, our approach involves a modified update scheme, which we prove possesses convergence guarantees that are similar to those for a traditional stochastic gradient (SG) method. We also present numerical results showing that TRish can outperform SG when solving convex and nonconvex machine learning test problems. n MC10 North Bldg 125A Online Platforms and Operations Management Sponsored: Manufacturing & Service Oper Mgmt Sponsored Session Chair: Eunae Yoo, University of Tennessee, 5416 W Cindy Place, Tempe, AZ, 85226, United States 1 - The Wisdom of Crowds in Operations: Forecasting using Prediction Markets Ruomeng Cui, Emory University, 1935 Ridgemont Lane, Decatur, GA, 30033, United States, Achal Bassamboo, Antonio Moreno Prediction is challenging when historical information is not available, such as forecasting demand of a new product. One approach that can be applied in such situations is to crowdsource opinions from employees and the public. We focus on a widely-adopted class of crowd-based forecasting tools—-prediction markets. These are virtual markets created to aggregate crowds’ opinions and operate in a way similar to stock markets. We partnered with Cultivate Labs. Using information extracted from employees and public crowds, we show that prediction markets produce well-calibrated forecasting results. In addition, we run several field experiments to study the conditions under which groups work well. 2 - Managing Service Systems in Presence of Social Networks Gad Allon, University of Pennsylvania, 3730 Walnut Street, Philadelphia, PA, 19104, United States, Dennis Zhang We study a service system with the presence of a social network. In our model, firms can differentiate resource allocations among customers, and customers learn the service qualities from the social network. We study the interplay among network structure, customer characteristics, and information structure, and characterize the optimal policy. We further calibrate our model with data from Yelp.com and quantify the value of social network knowledge empirically.

n MC08 North Bldg 124A Joint Session OPT/Practice Curated: Network Optimization in Power Systems Sponsored: Optimization/Network Optimization Sponsored Session Chair: Adolfo Raphael Escobedo, Arizona State University, Tempe, AZ, 85287-8809 1 - Optimal Portfolio of Power Flow Control Technologies: Topology and Impedance Control Mostafa Sahraei-Ardakani, University of Utah, 50 South Central Campus Drive, #2110 Merrill Engineering Building, Salt Lake City, UT, 84112, United States High congestion costs demand for more efficient utilization of the transmission system. Topology and impedance control are two technologies that enable such efficiency gains, through controlling the power flows. As the system operators begin to utilize these tools, it is essential to understand the interdependence between them. This talk presents simulation results that suggest a strong interdependence between topology and impedance control. It is, thus, essential to acknowledge this interdependence, both at the planning and operation stage. Failure to do so will lead to economic inefficiencies that can be avoided through appropriate co-optimization. 2 - Strategic Behavior of Self-scheduling Distributed Energy Resources in Energy and Reserve Co-optimizing Markets Fatma Selin Yanikara, Boston University, 15 St Marys Street, Room 140, Boston, MA, 02215, United States, Michael C. Caramanis We study the participation of distribution network connected, self-scheduling Distributed Energy Resources (DERs), such as EVs and wind farms, in energy and reserve markets. Important issues that arise when DERs self-schedule are existence and uniqueness of equilibrium, and strategic behavior. We study the distributed market clearing with self-scheduling DERs in game theory context and characterize the Nash equilibrium under various information availability cases: DERs are either (i)price takers or (ii)have access to local information and can explicitly calculate locational prices. Information aware DERs engage in a Nash game since they know how their own and others’ actions impact prices. 3 - Spatiotemporal Marginal Costs in Optimal Operational Planning of Distribution Networks Panagiotis Andrianesis, Boston University, Boston, MA, United States, Michael C. Caramanis The rapid growth of Distributed Energy Resources (DERs) presents a major challenge together with a still unexploited opportunity for a radical transformation of the distribution grid. We employ a distribution grid representation that captures the salient features and costs of distribution assets, as well as the complex DER preferences and capabilities, and enables the discovery of short-term dynamic locational marginal costs. We discuss the inherent difficulties of the operational planning optimization problem, and we present results from actual distribution feeders. 4 - Bus-angle Difference Valid Inequalities and Algorithms for DC Power Transmission Expansion Planning Kyle Skolfield, Arizona State University, Phoenix, AZ, United States, Laura M. Escobar, Adolfo Raphael Escobedo, Ruben Romero To meet rising demand for electricity under limited budgets, it is necessary to determine the best Transmission Expansion Planning strategies. This problem can be modeled as a large-scale MIP whose solution is intractable. To enable efficient search of the solution space, we derive a class of angular valid inequalities (AVIs) to be incorporated as cutting planes in the root node of the branch-and-bound tree. We design a data-driven scheme guided by solutions to various relaxation models to select the most effective AVIs. We test this scheme’s effectiveness via benchmark instances.

n MC09 North Bldg 124B Methods for Large Scale Nonlinear and Stochastic Optimization I Sponsored: Optimization/Nonlinear Programming Sponsored Session

Chair: Albert Solomon Berahas, Evanston, IL, 60208, United States Co-Chair: Francisco Jara-Moroni, Northwestern University, Evanston, IL, 60208, United States 1 - Gauss-newton Methods for Deep Neural Networks Yi Ren, Columbia University, 331 Mudd Bldg, I. E.O.R. Department, New York, NY, 10027-6699, United States,

187