2016 INFORMS Annual Meeting Program

TD66

INFORMS Nashville – 2016

3 - Facilitate Fit Revelation In a DistributionChannel Lin Hao, University of Notre Dame, 351 Mendoza College Of Business, Notre Dame, IN, 46556, United States, lhao@nd.edu, Yong Tan We investigate a retailer’s and a supplier’s incentive to facilitate fit revelation, i.e., facilitate consumer learning of their true product fit, under two popular channel pricing models, agency pricing model and wholesale pricing model. TD66 Mockingbird 2- Omni Technometrics Invited Session: Recent Statistical Chair: Peihua Qiu, Professor and Chair, University of Florida, 2004 Mowry Road, Gainesville, FL, 32611, United States, pqiu@phhp.ufl.edu 1 - Discovering The Nature Of Variation In Nonlinear Profile Data Daniel Apley, Northwestern University, apley@northwestern.edu Most prior work on profile data in the quality control literature has focused on monitoring to detect sudden changes in the characteristics of the profiles, relative to an in-control sample of profiles. In contrast, we develop an approach for exploratory analysis of a sample of profiles to discover the nature of any profile- to-profile variation present over the sample via manifold learning. Instead of analyzing parameter variation in some prespecified parametric profile model, our focus is on discovering and visualizing an appropriate characterization or parameterization of the variation, as a tool to facilitate discovery (and ultimately elimination) of its root causes. 2 - A Bootstrap Metropolis-hastings Algorithm For Bayesian Analysis Of Big Data Faming Liang, University of Florida, faliang@ufl.edu MCMC methods have proven to be a powerful tool for analyzing data of complex structures. However, their computer-intensive nature precludes their use for big data analysis. We propose the bootstrap Metropolis-Hastings (BMH) algorithm, which provides a general framework for how to tame powerful MCMC methods to be used for big data analysis; that is to replace the full data log-likelihood by a Monte Carlo average of the log-likelihoods that are calculated in parallel from multiple bootstrap samples. The BMH algorithm possesses an embarrassingly parallel structure and avoids repeated scans of the full dataset in iterations, and is thus feasible for big data problems. 3 - Online Updating Of Statistical Inference In The Big Data Setting Elizabeth D. Schifano, University of Connecticut, Storrs, CT, United States, elizabeth.schifano@uconn.edu, Ming-Hui Chen, Chun Wang, Jing Wu, Jun Yan, Yuping Zhang We present statistical methods for big data arising from online analytical processing, where data arrive in streams and require fast analysis without storage/access to historical data. In particular, we develop computationally efficient, minimally storage-intensive iterative estimating algorithms and statistical inferences for linear models and estimating equations that update as new data arrive. We propose goodness of fit tests, a new estimator within the estimating equation setting, and a modification to incorporate new variables midway through the data stream. We demonstrate the effectiveness of our procedures through theoretical and empirical analysis, as well as in application. TD67 Mockingbird 3- Omni QTQM Invited Session Sponsored: Quality, Statistics and Reliability Sponsored Session Chair: Jing Li, Arizona State University, Arizona State University, Tempe, AZ, 85287, United States, jing.li.8@asu.edu 1 - A Mixed-effect Model For Analyzing Experiments With Multistage Processes Kaibo Wang, Tsinghua University, kbwang@tsinghua.edu.cn In industrial practice, most products are produced by processes that involve multiple stages. In this paper, through an analysis of the error transmission mechanism, we propose a mixed-effect model for analyzing experiments with multistage processes. Based on an analysis of simulated and real experimental data, we find that different conclusions about factor significance may be drawn if the data are analyzed differently. In addition, the mixed-effect model can help separate errors at different stages and hence provide more information on process improvement. Methods for Analyzing Big Data Sponsored: Quality, Statistics and Reliability Sponsored Session

2 - Setup Adjustment For Asymmetric Cost Functions Under Unknown Process Parameters Arda Vanli, Florida State University, 2525 Pottsdamer St, Tallahassee, FL, FL, 32310, United States, oavanli@eng.fsu.edu, Zilong Lian, Enrique Del Castillo” We present a bayesian approach for the optimal control of a machine that can experience setup errors assuming an asymmetric off-target cost function. It is assumed that the setup error cannot be observed directly due to presence of measurement and part-to-part errors and the variance of this error is not known a priori. The setup error can be compensated by performing sequential adjustments of the process mean based on observations of the parts produced. We show how the proposed method converges to the optimal (known variance) trajectory, recovering from a possibly biased initial variance estimate. Simulations results are presented for constant asymmetric and quadratic asymmetric cost functions. 3 - Bounded Loss Functions And The Characteristic Function Inversion Method For Computing Expected Loss Matthias Tan, City University of Hong Kong, matthtan@cityu.edu.hk In robust parameter design, the quadratic loss function is commonly used. However, this loss function is not always realistic. We propose a general class of bounded loss functions that are cumulative distribution functions and probability density functions. New loss functions are investigated and they are shown to behave differently from the quadratic loss. For models linear in the noise factors, we give a numerical method based on characteristic functions inversion for computing expected loss. The method is very quick and accurate. It is applicable as long as the distributions chosen to represent the loss function and variation in the noise factors have tractable characteristic functions. 4 - Quasi-feedforward And Feedback Control For Random Step Shift Disturbance Models Lihui Shi, Senior Data Scientist, Centerfield Corporation, El Segundo, CA, 90245, United States, shilihui@uw.edu Lihui Shi, Senior Data Scientist, University of Washington, Seattle, WA, 98195, United States, shilihui@uw.edu Process monitoring and process adjustment strategies are two important parts of the process improvement methods, and they should be integrated together in stead of separated. Integrated moving average (IMA) model is the most common disturbance model, and step shift model is one type of more complicated one that often exists in real applications. We investigate the IMA background disturbance subject to random step shifts with a certain probability. We propose a feedback control with a quasi-feedforward control by monitoring the output errors. It is very robust against parameter misspecifications. We also investigate the type I and type II errors in process adjustment on this disturbance model. TD68 Mockingbird 4- Omni Reliability Modeling and Optimization in Early Product Design Stages Sponsored: Quality, Statistics and Reliability Sponsored Session Chair: Zhaojun Li, Western New England University, Springfield, MA, United States, zhaojun.li@wne.edu 1 - Assessing Failure Dependency In A Complex System Rong Pan, Arizona State University, Rong.Pan@asu.edu, Petek Yontay In this talk we will discuss a Bayesian network model for assessing system reliability of a complex system. Coupling with Bayesian inference methods , the posterior distributions of the conditional probabilities in a BN model can be estimated by combining failure information and expert opinions at both system and component levels. 2 - A Multi-objective Approach For Multi-stage Reliability Growth Planning By Considering The Timing Of New Technologies Introduction Steven LI, Western New England University, zhaojun.li@wne.edu, Mohammad Sadegh Mobin, Hossein Cheraghi This paper proposes a new multi-stage reliability growth planning model which optimizes and balances the product development cost, time, and the product reliability. The number of test units, test time, and the percentage of introduced new technologies are major decision variables. Different reliability growth rates are considered for each sub-system in each stage. An integrated approach is developed to optimize the problem, which starts with a multi-objective evolutionary algorithm to find a set of Pareto optimal solutions followed by the application of clustering tools to cluster the solutions. The clustered solutions are further ranked using a multiple criteria decision making tool.

356