Informs Annual Meeting Phoenix 2018

INFORMS Phoenix – 2018

SD10

n SD08 North Bldg 124A Joint Session OPT/Practice Curated: Network Design Models and Applications Sponsored: Optimization/Network Optimization Sponsored Session Chair: Navid Matin Moghaddam, Arizona State University, AZ 1 - Large-scale Network Design with Application to Carbon Capture and Storage Navid Matin Moghaddam, Tempe, AZ, 85281, United States, Jorge A. Sefair Motivated by the emerging need of climate stabilization, CCS seeks to prevent carbon emissions from entering the atmosphere by capturing, compressing, and transporting CO2 from sources (e.g., oil refineries) to geologic reservoirs (e.g., depleted oil wells) for its long-term storage. To find the least-cost network of pipelines able to transport a target amount of CO2, we propose an optimization model and an exact decomposition algorithm that takes advantage of the optimal solution sparsity and the problem’s cost structure. We illustrate the performance of our methods on a set of real instances. 2 - Profit Oriented Fixed-charge Network Design with Elastic Demand Carlos Armando Zetina, Concordia University, 1455 de Maissonneuve Boulevard West, Montreal, QC, H3G 1M8, Canada, Ivan Contreras, Jean-Francois Cordeau Networks require significant investments and their infrastructure is built to last over an extended period of time. It is unrealistic to believe that the demand patterns will remain the same during its lifetime. Moreover, demand patterns depend on the resulting network’s topology, leading to sub-optimal designs if static estimations are used. We attempt to circumvent this by incorporating elastic demand into a fundamental network design problem. We replace static origin- destination demand estimations with functions of the network characteristics often used to estimate demand. This provides the decision-maker with insights on the effect the design has on demand patterns. 3 - Using Network Optimization to Extend the Transit Opportunity Index for Large-scale Transit Systems Andres L. Medaglia, Professor, Universidad de los Andes, Cr 1 este Bogota, 111711, Colombia, Claudia D. Bedoya, Olga L. Sarmiento, Daniel A. Rodriguez, Jessica Camacho The Transit Opportunity Index (TOI) quantifies accessibility and connectivity of a public mass transit system by combining spatial, temporal, and trip coverage metrics. We extended the TOI to allow for connectivity in large-scale transit systems where more than one transfer may be required. In this extended TOI we build a time-space network from the public transit supply and find shortest paths allowing multiple transfers. With the extended TOI, transit planners could adjust transit service to include opportunities and accessibility as design goals and address equity by identifying underserved areas. We illustrate the extended TOI with the public transit network of Bogotß, Colombia. 4 - The Wireless Network Design Problem Subject to Radio Interference Hugh Medal, Mississippi State University, P.O. Box 9542, Msu, MS, 39762, United States, Will Leonard, Randy K. Buchanan This paper studies the wireless network design problem, which considers how best to place a set of antennas so that the antennas can send and receive data in a way that maximizes total network throughput. A mixed-integer program was implemented to solve this problem using a Benders-decomposition approach. Maximal independent sets representing antennas’ connections to one another were produced dynamically. The goal of this paper is to demonstrate whether utilizing directional antennas instead of omnidirectional antennas affects maximum throughput, as well as what changes occur to maximum throughput based on imposing constraints related to power and channel availability.

n SD09 North Bldg 124B Intersection between Optimization and Statistical Learning Sponsored: Optimization/Nonlinear Programming Sponsored Session

Chair: Hongcheng Liu, University of Florida, Gainesville, FL, 1 - Global Optimization of High-dimensional Generalized Linear Models with Nonconvex Regularization in Polynomial-time with High Probability Charles Hernandez, University of Florida, Gainesville, FL, 32606, United States In this presentation we investigate a local optimization scheme for a class of nonconvex learning problems known to be NP-Hard to solve. This optimization scheme can be shown to find a global optimum in polynomial time with high probability. The approach uses LLA in combination with a Folded Concave Penalty function and can be applied to the training of many high-dimensional generalized linear models. 2 - Second-order Condition in Nonsmooth High-dimensional Learning Hongcheng Liu, University of Florida, Weil Hall 478, Gainesville, FL, 32611, United States This research concerns the high-dimensional learning with non-smooth empirical loss functions. Through the use of a folded concave penalty (FCP), we show that the excess risk of the M-estimation problem can be well contained even if the loss function is not necessarily differentiable and the number of dimensions cannot be bounded by any polynomial function of the sample size. Such a desirable sample complexity is provably achieved at a second-order stationary point of the FCP- based nonconvex formulation. 3 - Linearly-constrained High Dimensional Learning Hung Yi Lee, University of Florida, 303 Weil Hall, Gainesville, FL, 32603, United States This presentation is mainly concerned with the computational and statistical complexities involved in the folded concave penalized high-dimensional M- estimation problems with linear constraints. We have shown that by using some local solutions which satisfy certain second-order necessary optimality conditions, the required sample size grows in poly-logarithmic of the number of dimensions, hence, improving the statistical performance. 4 - Complexity for Training ReLU Neural Network Digvijay Boob, Georgia Institute of Technology, Atlanta, GA, United States, Santanu Subhas Dey, Guanghui Lan A neural network is a function computed on a graph parameterized by edge weights. Problem of training neural network can be thought of as searching for edge weights such that output of neural network exactly matches with output according to data. It is known that training binary neural network is NP-complete whereas ReLU neural network is employed extensively in practice. Training algorithm involves heuristic methods such as stochastic gradient descent. There is a no complexity based hardness results for fully connected ReLU network. We show that training a simple fully connected ReLU neural network is NP-hard. On the other hand, applying over-prametrization to these networks render the problem in P. n SD10 North Bldg 125A Resource-Constrained Scheduling in Production and Project Management Sponsored: Manufacturing & Service Oper Mgmt Sponsored Session Chair: Norbert Trautmann, University of Bern, FM Quantitative Methoden, Schuetzenmattstrasse 14, Bern, 3012, Switzerland 1 - The RCPSP: New Benchmark Results Stefan Creemers, KU Leuven, Naamsestraat 69, Leuven, 3000, Belgium, Stefan Creemers, IESEG School of Management, Lille, France We present a new state-of-the-art procedure for solving the resource-constrained project scheduling problem. We present new benchmark results, and are able to solve several problem instances that were previously unsolved.

93