Informs Annual Meeting Phoenix 2018

INFORMS Phoenix – 2018

SA02

n SA02 North Bldg 121B Joint Session OPT-Uncert/APS: Optimization in Statistics I Sponsored: Optimization/Optimization Under Uncertainty Sponsored Session Chair: Robert Bassett, Naval Postgraduate School, Monterey, CA, United States 1 - Variational Analysis of M-estimators Johannes Royset, Naval Postgraduate School, Operations Research Department, Monterey, CA, 93943, United States We present a framework for construction, analysis, and computations of M- estimators in the presence of soft information about shape, support, continuity, slope, location of modes, density values, and other conditions that, individually or in combination, restrict the family of estimates under consideration. The framework also leads to plug-in estimators with the exceptional property that convergence of densities and regression functions implies convergence of modes, maximizers, high-likelihood events, and related quantities. We illustrate the approach with several applications. 2 - Approximation of Maximum a Posteriori Estimators Using Bayes Estimators Julio Alejandro Deride Silva, Sandia National Laboratories, CERL/111, MS/1327, P.O. Box 5580, Albuquerque, NM, 87185- 1327, United States, Robert Bassett In this talk, we present an approximation scheme for statistical estimation using variational analysis techniques. In particular, we focus on the folk theorem that characterizes Maximum a posteriori estimators as a limit of a sequence of Bayes estimators under 0-1 loss. We first provide a counterexample, which shows that in general, this claim fails to hold true. Secondly, since both estimators are defined in terms of optimization problems, the tools of variational analysis find a natural application to Bayesian point estimation, and we analyze the convergence using notions of hypo-convergence. Finally, we provide conditions when this result holds true 3 - Coupled Learning Enabled Optimization Junyi Liu, University of Southern California, 30 Virginia Avenue, Los Angeles, CA, 91107, United States, Guangyu Li, Suvrajeet Sen Empowered by machine learning techniques, diagnostic and predictive analytics are usually followed by decision-making problems in prescriptive analytics. We extend the above sequential prediction-optimization paradigm to a case in which prediction and optimization model are coupled so that the parameter estimation model can guide the optimization models to achieve higher levels of performance. Under certain assumptions on the problem and training data set, we show that the sequence of solutions provided by the coupled algorithm converges to the first-order stationary point of the original stochastic optimization problem. 4 - Phase Retrieval Denoising via Rapid Eigenvalue Computation of Evolving Matrices Will Wright, University of California, Davis, CA, United States, Zhaojun Bai Phase retrieval has a wide range of solution methods, yet few exist for handling noisy observations (aka, phase retrieval denoising) without imposing additional restrictions such as signal sparsity. We focus on a recent phase retrieval denoising model: the gauge dual of the PhaseLift model (GD-PL). GD-PL can be solved with subgradient descent, with the main cost being a sequence of evolving eigenvalue problems. We establish optimal methods for handling this sequence of eigenvalue problems. We also establish probabilistic results of the optimality of the signal returned by the subgradient descent algorithm. 5 - Fused Density Estimation for Data on Infrastructure Networks Robert Bassett, Naval Postgraduate School, 1 University Circle, Monterey, CA, United States, James Sharpnack We introduce a new method for nonparametric density estimation on geometric networks. By penalizing maximum likelihood estimation with a total variation penalty, we avoid overfitting and the dirac curse. We provide results which reduce the search space for the estimator from infinite dimensional function space to the finite-dimensional setting, and further demonstrate its computational tractability. We then focus on the asymptotic convergence rate of this density estimation method. Lastly, we review applications to infrastructure networks.

n SA03 North Bldg 121C Creating Value in Innovation Sponsored: Technology, Innovation Management

& Entrepreneurship Sponsored Session Chair: Tian Chan, Emory University’s Goizueta Business School, Atlanta, GA, 30322, United States 1 - The Firm Productivity Implications of Technology Licensing: Evidence from Developing Asian Economy Manufacturing Firms Xiaojin Liu, 100 Darden Boulevard, Charlottesville, VA, 22903, United States, Anant Mishra This study focuses on the productivity implication of technology licensing in developing Asian economy manufacturing firms, and examines how it is affected by infrastructural factors in a firm’s internal and external environment. 2 - Delegating Innovation Morvarid Rahmani, Georgia Institute of Technology, 800 West Peachtree Street NW, Room 4246, Atlanta, GA, 30308-1149, United States, Karthik Ramachandran In many contexts such as product design and advertising, clients seek the expertise of external providers to generate innovative solutions for their business problems. In this paper, we explore how the client’s flexibility in terminating the project can influence the progress and efficiency of the delegated innovation. 3 - Search and Sequential Innovation in Mobile App Development Nilam Kaushik, University College London, London, 02115, United Kingdom, Bilal Gokpinar The process of search, identification, and acquisition of knowledge is essential for the success of products. This paper empirically characterizes the sequential innovation process in the setting of mobile app development using novel text- mining techniques. 4 - Designing Internal Innovation Contests Lakshminarayana Nittala, University of California-San Diego, Rady School of Management, 9500 Gilman Drive, La Jolla, CA, 92093, United States, Sanjiv Erat, Vish Krishnan Firms can use internal contests to source solutions to problems associated with innovation. However, designing such contests involves nuanced understanding of the impact of such contests on the on-going projects within the firm. Optimal contest design is discussed along with managerial implications. n SA04 North Bldg 122A Joint Session OPT/Practice Curated: Theory and Applications of MIP Sponsored: Optimization/Integer and Discrete Optimization Sponsored Session Chair: Haochen Luo, Texas A&M University, College Station, TX, 77843-3131, United States 1 - A Bounded Formulation for the School Bus Scheduling Problem Liwei Zeng, Northwestern University, 2145 Sheridan Road, IEMS Department, Evanston, IL, 60208, United States, Karen Smilowitz, Sunil Chopra This paper proposes a new formulation for the school bus scheduling problem (SBSP) which optimizes starting times for schools and associated bus routes to minimize transportation cost. Specifically, the problem determines the minimum number of buses required to complete all bus routes under the constraint that routes for the same school must arrive within a set time window before that school starts. We present a new integer linear programming (ILP) formulation for this problem which is based on a time-indexed formulation. We develop a randomized rounding algorithm based on the linear relaxation of the ILP that yields near-optimal solutions for large-scale problem instances. 2 - Integer Programming for Learning Directed Acyclic Graphs Hasan Manzour, University of Washington, Seattle, WA, United States, Simge K. Á. Kyavuz, Ali Shojaie Bayesian Networks (BNs) are probabilistic graphical models that represent causal relationships among a set of random variables in the form of a Directed Acyclic Graph (DAG). We study the problem of DAG structural learning of a BN from observational data where the underlying causal mechanism in the network is linear. We propose a new optimization model for this learning problem and discuss the statistical implications of L1 versus L0 penalty in our model. The computational results, tested on both synthetic and real datasets, demonstrate that the proposed model is computationally more efficient to learn the optimal DAG when compared with the existing mathematical models in the literature.

2