Informs Annual Meeting Phoenix 2018

INFORMS Phoenix – 2018

TB69

4 - Wavelet-based Profile Monitoring using Order-thresholding Recursive CUSUM Schemes Ruizhi Zhang, Georgia Institute of Technology, Atlanta, GA, United States, Yajun Mei, Jianjun Shi With the rapid development of advanced sensing technologies, rich and complexreal-time profile or curve data are available in many manufacturing processes. These profile data provide valuable intrinsic information about the performance of the process, subject or product, and it is often desirable to utilize them to develop efficient methodologies for process monitoring and fault diagnosing. In this article, we propose a novel wavelet-based profile monitoring procedure that is based on the order-thresholding transformation of recursive CUSUM statistics of multiple wavelet coefficients. Simulation and case study are conducted to illustrate the usefulness of our proposed procedure. n TB69 West Bldg 106A Large Scale Statistical Learning General Session Chair: Meisam Razaviyayn, University of Southern California, Los Angeles, CA, 90089, United States 1 - Learning Bounds for Greedy Approximation with Multiple Explicit Feature Maps Shahin Shahrampour, Texas A&M University, College Station, TX, United States Large-scale Mercer kernels can be approximated using low-dimensional feature maps for efficient risk minimization. Due to the inherent trade-off between the feature map sparsity and the approximation accuracy, the key problem is to identify promising feature map components (bases) leading to satisfactory out-of- sample performance. In this work, we tackle this problem by efficiently choosing such bases from multiple kernels in a greedy fashion. Our method sequentially selects these bases from a set of candidate bases using a correlation metric. We prove that the out-of-sample performance depends on three types of errors, one of which (spectral error) relates to spectral properties of the best model in the Hilbert space associated to the combined kernel. The result verifies that when the underlying data model is sparse enough, i.e., the spectral error is negligible, one can control the test error with a small number of bases, scaling poly- logarithmically with data. Our empirical results show that given a fixed number of bases, the method can achieve a lower test error with a smaller time cost, compared to the state-of-the-art in data-dependent random features. 2 - Valid p-values from Adaptively Collected Data Yash Deshpande, Massachusetts Institute of Technology, Cambridge, MA, United States Data collection in many scientific and engineering applications is inherently adaptive. For instance, sequential or multi-stage clinical trials may reallocate patients among different treatments based on data gleaned from previous stages. In e-commerce applications, most user behavior data is collected under feedback from recommendation algorithms that are continuously trained. Such adaptivity is problematic for ex post inference, causing standard statistical techniques to fail. We propose a novel decorrelating procedure that correct standard estimators for adaptivity. The procedure is model-free: it requires only coarse information about the data-collection algorithm, in comparison with competing procedures like propensity scores. Therefore, it enables computing ex post valid p-values and confidence intervals on adaptively collected data. 3 - On the Importance of Being Improper in ML Karthik Sridharan, Cornell University, Ithaca, NY, United States This talk will focus on the benefits of improper learning (where predictor can lie outside model class) when dealing with machine learning problems. Be it from the perspective of computational efficiency, sample complexity or both. The first part of the talk with focus on sample complexity. We will specifically focus on logistic regression and Starting with the simple observation that the logistic loss is 1-mixable, we design a new efficient improper learning algorithm for online logistic regression that circumvents the aforementioned lower bound with a regret bound exhibiting a doubly-exponential improvement in dependence on the predictor norm. This provides a positive resolution to a variant of the COLT 2012 open problem of McMahan and Streeter (2012) when improper learning is allowed. The second part of the talk will focus on computational advantages of improper learning. 4 - Randomized Algorithms for Infinite Dimensional Optimization Problems Cameron Musco, Massachusetts Insititute of Technology, Cambridge, MA, United States We will discuss recent advances in randomized algorithms for infinite dimensional optimization problems that arise in kernel-based machine learning. Kernel-based learning is notoriously difficult to scale to large datasets because the runtime of exact methods typically scales quadratically in the number of data points. Many techniques have been suggested to reduce this cost, including Rahimi and Recht’s random Fourier features method and the Nystr÷m method, which both rely on random sampling to obtain approximately optimal solutions.

n TB70 West Bldg 106B Joint Session DEA/Practice Curated: Theoretical issues in DEA Emerging Topic: Productivity, Efficiency and Data Envelopment Analysis Emerging Topic Session Chair: Peter Bogetoft, Frederiksberg, 2000, Denmark 1 - Rate Design and Grid Efficiency Jung Sook You, Assistant Professor, California State University- East Bay, 25800 Carlos Bee Blvd, Hayward, CA, 94542, United States Time-dependent consumer preferences combined with intermittent renewable energy sources may result in excess demand fluctuations on a power grid network. Assuming consumers and generators can be price-anticipating strategic players, this paper derives the optimal locational marginal prices on the power grid with conventional and renewable generators and provides theoretical prediction of price volatility. Then, the paper simulates the extent to which the optimal locational prices respond to the volatility of renewable sources and consumer demands, using California wholesale market data. 2 - Benchmarking, Contracting, and Incentive Power Peter Bogetoft, Copenhagen Business School, Porcelaenshaven 16 A, Frederiksberg, 2000, Denmark Benchmarking is a popular management tool. It is used both to facilitate decision making and incentive provision. In this paper, we show that different benchmarking measures react differently to changes in performance. Therefore, the optimal design of benchmarking based performance contract requires an explicit analysis of the data generation process of the benchmarking approach. As an illustrative application, we investigate the use of Data Envelopment Analysis (DEA), Stochastic Frontier Analysis (SFA) and combinations hereof in revenue cap incentive schemes. Such schemes are applied in the regulation of network companies in many countries. Chair: Anna B. Nagurney, University of Massachusetts-Amherst, Isenberg School of Management, Dept of Operations & Information Mgmt, Amherst, MA, 01003, United States 1- Fortification Against Cascade Propagation Under Uncertainty Colin P. Gillen, University of Pittsburgh, 1025 Benedum Hall, 3700 O’Hara Street, Pittsburgh, PA, 15261, United States Network cascades represent a number of real-life applications: social influence, electrical grid failures, viral spread, etc. The commonality between these phenomena is that they begin from a set of seed nodes and spread to other regions of the network. We consider a critical elements detection problem where the decision-maker wishes to limit cascading behavior via node threshold modification (within a budget). The arc weights - how much influence one node has on another - are uncertain. The solution approaches include robust mixed- integer programming, an expand-and-cut algorithm, and centrality-based heuristics. Extensive computational experiments demonstrate the value of these methods. 2 - V-Polyhedral Disjunctive Cuts Aleksandr M. Kazachkov, Carnegie Mellon University, Tepper School of Business, 5000 Forbes Avenue, Pittsburgh, PA, 15213-3890, United States We introduce V-polyhedral cuts (VPCs) for generating strong disjunctive inequalities, with the goal of mitigating numerical issues in existing cuts by avoiding recursion. The VPC framework involves collecting a properly selected set of points and rays. This affords computational advantages over existing methods and enables testing strong multiterm disjunctions arising from partial branch-and- bound trees. Our computational and theoretical results suggest the potential practical benefits of VPCs, while motivating future work on better understanding the interaction between branching and cutting. n TB71 West Bldg 106C ICS Student Paper Prize Sponsored: Computing Sponsored Session

300