|
2003 Quality & Productivity Research ConferenceIBM T. J. Watson Research Ctr., Yorktown Heights, NY May 21-23, 2003 |
Contributed Paper Sessions (with Abstracts and Papers)
1. Industrial Applications
Session Chair: Andre Pinho, University of Wisconsin
1. "The Six Sigma Approach to Reduce the Product Damages in the Warehouse",
Nihal M. Erginel, Anadolu University, Turkey, and Nimetullah Burnak, Osmangazi
University, Turkey. Paper
Abstract: The most common complaint of end customers about the defects is the product
damages to the warehouse in which the products are packaged, stored, and
shipped. Six sigma methodology is applied to eliminate the complaints about
the product damages. The factors which are assumed to be effective on the
occurrence of the damages are determined by using process map, and cause
ö effect diagram. Their priorities are assigned via cause ö effect matrix.
It is decided that the product damages can occur in any one of the two
phases. The first phase is the packaging for which packaging materials
and packaging method are considered. The second phase is the transportation
for which handling equipments are analyzed.
DOE is conducted separately for each of the phases, and the results are discussed.
2. "Improvement of Heating Component Production Process Using 6 Sigma Methodology", Berna ATA, Korel Elektronik, Eskisehir, Turkey, and Nimetullah Burnak, Osmangazi University, Turkey.
Abstract: Product improvement, productive manufacturing are two of the important
topics of the companies in order to compete. Companies must consider productivity
and efficiency. Under severe competitive conditions, in order to increase
the market share requires to decrease or eliminate variation in the production
process and improve product quality. 6 sigma methodology is a useful tool
to improve the production process. In this study, the improvement studies
with 6 sigma of a heating component used to defrost in a cooling product
are discussed. Starting from problem definition to the results obtained
such as type of fiberglass, diameter of the wire, etc. are covered.
3. "Assessing Quality of Asphalt Paving Jobs to Determine Contractor
Pay", Robin C. Wurl and James R. Lundy, Oregon State University. Paper
Abstract: A quality assessment procedure for asphalt mix production is developed
for the Oregon Department of Transportation to link contractor pay to expected
field performance. The methodology, based on a loss function, encourages
the production of asphalt mixes that are consistent with specifications
with minimum variability. The mix has multiple quality characteristics
but not all are equally important and the specifications may change during
the lifetime of a job.
4. "Estimating the efficiency of collaborative problem-solving, with
applications to chip design", Mary Y. Lanzerotti - Wisniewski, E.
Yashchin, R.L. Franch, D.P. Conrady, G. Fiorenza, I.C. Noyan, IBM Research
Paper
Abstract: We present a statistical framework to address questions that arise in
general problems involving collaboration of several contributors. One instance
of this problem occurs in the complex process of designing ultralarge-scale-integration
semiconductor chips. In cases involving complex designs, the computer-aided
design tools are unable to create designs that satisfy specified project
criteria, and a number of questions arise about how to measure the effectiveness
of systematic external intervention that is implemented with some supplemental
algorithm. As an example, we apply the statistical framework to the problem
of routing a functional unit of the IBM POWER4 microprocessor.
2. Process Control I
Session Chair: Norma Leyva - Estrada, Iowa State University
5. "Using Statistical Process Control To Monitor Active Managers",
Thomas K. Philips, Paradigm Asset Management, Emmanuel Yashchin, IBM Research,
David M. Stein, Parametric Portfolio Associates Paper
Abstract: Investors who are invested in (or bear responsibility for) many active
portfolios face a resource allocation problem: To which products should
they direct their attention and scrutiny? Ideally they will focus their
attention on portfolios that appear to be in trouble, but these are not
easily identified using classical methods of performance evaluation. In
fact, it is often claimed that it takes forty years to determine whether
an active portfolio outperforms its benchmark. The claim is fallacious.
In this article, we show how a CUSUM process control scheme can be used
to reliably detect flat-to-the-benchmark performance in forty months, and
underperformance faster still. By rapidly detecting underperformance, the
CUSUM allows investors to focus their attention on potential problems before
they have a serious impact on the performance of the overall portfolio.
The CUSUM procedure proved to be robust to the distribution of excess returns,
allowing its use in almost any asset class, including equities, fixed income,
currencies and hedge funds without modification, and is currently being
used to monitor over $500 billion in actively managed assets.
6. "Phase I Monitoring of Nonlinear Profiles", James D. Williams,
William H. Woodall and Jeffrey B. Birch, Virginia Polytechnic Institute
& State University. Paper
Abstract: In many quality control applications, a single measurement is insufficient
to characterize the quality of a produced item. In an increasing number
of cases, a profile (or signature) is required, often consisting of several
measurements of the item across time or space. Such profiles can frequently
be modeled using linear regression models or a nonlinear regression models.
In recent research others have developed multivariate T2 control charts
for monitoring the coefficients in a simple linear regression model of
a profile. However, little work has been done to address the monitoring
of profiles that can be represented by a parametric nonlinear regression
model. Here we extend the use of the T2 control chart to monitor the coefficients
of the nonlinear regression fits to the profile data. We give several general
approaches to the formulation of the T2 statistics and the associated upper
control limits in Phase I applications. Finally, these approaches are illustrated
using the vertical board density profile data used by Walker and Wright
(JQT, 2002).
7. "On the design of Bayesian Control Charts using Markov Decision Processes", Bianca M. Colosimo, Politecnico di Milano (Italy)
Abstract: Following the first pioneristic work by Girshick and Rubin (1952), different
studies showed that traditional non-Bayesian control charts are not optimal
from an economic point of view. Despite of the "significant theoretical
value of Girshick and Rubin's model" (Montgomery, 2001), this approach
has had little attention because of the computational complexity required
to derive the optimal control rule.
During the last decade, Bayesian control charts received a renewed attention in the framework of Adaptive Control Charts (Tagaras 1994, 1996, 1998), and dynamic programming was presented as the viable solution to overcome computational difficulties. Following the work of Smallwood and Sondik (1973), this paper re-address the problem of designing a Bayesian control chart using newly developed algorithms developed in the framework of Partially Observable Markov Decision Processes (POMDP).
3. Design of Experiments
Session Chair: Alexandra Kapatou, University of Michigan
9. "Construction of Optimal Constrained Permutation Mixture Experiment
Designs", Ben Torsney and Yousif A. Jaha, University of Glasgow. Paper
Abstract: We construct 'D-optimal' constrained permutation mixture exper-iment
designs. Design points are permutations of a single set of (non-negative)
proportions summing to 1, thereby meeting conditions satisfied by 'mixture'
variables. Two types of constraint are considered on the proportions :
order constraints (for finding local maxima); common lower and/or upper
bound constraints. In both cases the constrained optimisation problem can
be transformed to one in which optimisation is with respect to a new set
of proportions or convex weights. A multiplicative algorithm is used to
optimise the D-criteria over the proportions under a Scheffe model. Results
extend to blocking and other models.
10. "New Results about Randomization and Split-Plotting", James M Lucas. Paper
Abstract: New results include this fact that the Kiefer-Wolfowitz equivalence theoremdoes
not hold for split-plot experiments. D- and G- criterion give different
designs. Computer approaches to design must recognize this. New design
examples will be given. This is joint work with Peter Goos.
I would also discuss "SUPER-EFFICIENT" experiments and give catalogues
of optimum blocking for one and two hard-to-change factors. This is joint
work with Frank Anbari, Derek Webb and John Borkowski (who is attending
and presenting designs generated using a genetic algorithm). Examples of
Biases form running RNR (Randomized Not Reset) experiments (when the factor
is not set to a neutral level and then reset when successive runs have
the same level) will be presented. This is joint work with Jeetu Ganju.
11. "Post-Fractionated Strip-Block Designs: A Tool for Robustness
Applications and Multistage Processes", Carla A. Vivacqua, University
of Wisconsin, S¿ren Bisgaard, University of Massachussets, Harold J. Steudel,
University of Wisconsin Paper
Abstract: This paper presents a novel experimental arrangement, called post-fractionated
strip-block design, which represents a cost-effective method to gather
knowledge and fast responses for guiding the design of robust products
while reducing product development expenses. It can also be applied in
the improvement of multistage processes and in studies involving hard-to-change
factors.
12. "Computer Experiments: Designs to Achieve Multiple Objectives",
Leslie M. Moore, Los Alamos National Laboratory
Abstract: Orthogonal arrays, or highly fractionated factorial designs, are suggested
for computer experiments in which goals may include sensitivity analyses
or response surface modeling. Latin hypercube samples, possibly selected
by space-filling criterion, are commonly used when Gaussian spatial processes
are the modeling paradigm of choice or uncertainty analysis is the objective.
Designs with more than 2 or 3 levels per input or densely covered 1 or
2 dimensional projections are also desirable. Competing experiment objectives
will be discussed and experiments are suggested that combine designs with
different properties.
4. Process Control II
Session Chair: Daniel R. Jeske, Lucent
13. "Control Charts for Binomial Proportions", John Aleong, University of Vermont
Abstract: The problem of the control chart for the binomial proportion (p-charts)
will be revisited. The p-chart is based on the standard Wald confidence
interval. The erratic behavior of the coverage probability of the Wald
confidence interval is discussed by Blyth & Still(1983), Agresti and
Coull(1998), Santner(1998) and others. Recent results by Brown, Cai, and
DasGupta (2002, 2001,2000), have shown theoretically, using Edgeworth expansions
and simulations the eccentric behavior of the Wald confidence interval
for various values of p and n. Using the coverage probabilities and expected
length of the intervals, they compared the Wald confidence interval with
other intervals. Brown et al recommended the Wilson intervals (Wilson 1927)
and the Jeffrey prior interval for small n, while for large n the Agresti
and Coull(1998). These results have consequences for the p-charts in current
use. P öcharts based on the results of Brown et al will be presented and
illustrated with recommendations.
14. "A Transition Matrix Representation of The Algorithmic Statistical Process Control Procedure with Bounded Adjustments and Monitoring", Changsoon Park, Chung-Ang University Paper
Abstract: In processes where both an irremovable disturbance and a special cause
are assumed to exist, procedures for adjustment and monitoring are necessary
for controlling the process level close to target. Such a combined procedure
for adjustment and monitoring is termed as an algorithmic statistical process
control (ASPC) procedure.
A transition matrix representation is developed to derive the properties of the ASPC procedure under an IMA(0,1,1) disturbance model. States of the transition matrix are constructed when the ranges of the two control statistics, one is the predicted deviation and the other is the EWMA chart statistic, are classified into a certain number of subintervals, and values in each interval are represented by a discrete value. Then the properties of the ASPC procedure are derived by the operation of the transition matrix. Each element of the transition matrix, i.e. the transition probability from a prior state to a posterior state, is calculated according to the given states and the process level.
This technique can be easily applied to the ASPC procedure with repeated adjustments instead of bounded adjustments.
15. "Statistical Quality Control Techniques for Low Volume and Short-Run Production", Tyler Mangin and Canan Bilen, North Dakota State University Paper
Abstract: The trend in advanced manufacturing industries has been to shift from
mass production to low volume production in order to meet customer demand
for smaller, more frequent deliveries. This has resulted in a need to develop
statistical quality control techniques that are effective in low volume
manufacturing environments. These techniques must address critical issues
specific to the low volume environment including: charting multiple product-types
processed on a single machine, frequency and effect of machine setups on
process variability, data scarcity, and process adjustment decisions. The
application of existing control charting methodologies in a low volume
production environment will be addressed, with emphasis on electronics
assemble and manufacturing.
5. Statistical Methods I
Session Chair: J.D. Williams, Virginia Tech
16. "Bayesian Inference for PVF Frailty Models ", Madhuja Mallick
and Nalini Ravishanker, University of Connecticut Paper
Abstract: In this article, we describe inference for multivariate lifetimes data using a conditional proportional hazards model with a power variance family (PVF) frailty distribution and a piecewise exponential hazard with correlated prior process baseline hazard. The likelihood function is derived as the joint density of tilted positive stable random variables. Inference is carried out in the Bayesian framework, using Markov chain Monte Carlo techniques. We illustrate our approach on data involving recurrent infections due to insertion of a catheter in patients on portable dialysis machines.
18. "Analysis of Repairable Systems: MTBF Versus MCF", David
Trinidade, Sun Microsystems.
6. Statistical Methods II
Session Chair: Carla Vivacqua, University of Wisconsin
19. "Estimating Sensitivity Of Process Capability Modeled By A Transfer
Function", Alan Bowman and Josef Schmee, Graduate Management Institute,
Union College
Abstract: Assume that the output variable Y of a process with a number of inputs
(X1, X2,..., Xn) is subject to specification limits and that the inputs
can be represented as random variables. The probability that Y falls within
its specification limits is a measure of process capability. In this paper,
we demonstrate a one-pass Monte Carlo simulation method that allows the
estimation of the sensitivity of the process capability to each parameter
of the input variables. The results can be used in improving system performance
by directing the analyst to those parameters for which small changes result
in the largest change in capability. The paper outlines the algorithm,
demonstrates it on three problems and provides some intuition as to why
it works.
20. "Follow-up Experiments to Remove Confounding Between Location and Dispersion Effects in Unreplicated Two-Level Factorial Designs", AndrŽ L. S. de Pinho, University of Wisconsin, S¿ren Bisgaard, University of Massachussets, Harold J. Steudel, University of Wisconsin Paper
Abstract: The objective of this paper is to present a methodology that allows us
to select the minimum necessary number of trials to gather new information
to help resolve the ambiguity among concurrent models. We extend the discrimination
criterion of Meyer, Steinberg and Box (1996) by allowing a non-homogeneous
variance scenario. The results are exemplified on Montgomeryâs (1990) injection
molding experiment.
21. "Two New Mixture Models for Living with Collinearity but Removing
its Influence ", John A. Cornell, University of Florida
Abstract: Fitting equations to mixture data collected from highly constrained regions
has challenged modelers for the past 35 years. Collinearity among the terms
in the models resulting in imprecise coefficient estimates is one of the
problems encountered. Recently, two new model forms have been introduced
where the terms are scaled and thus remove the influence of collinearity.
The benefits of fitting the new models is illustrated using two numerical
examples.
22. "Analyzing Supersaturated Designs Using Biased Estimation", Adnan Bashir and James R. Simpson, Florida State University Paper
Abstract: A designed experiment which investigates m number of factors with n number
of runs, where m>n-1 is referred to a supersaturated design. Supersaturated
designs recently received increased attention due to their use in investigating
many factors with fewer experiments. Stepwise regression is the most widely
used method to analyze the data of supersaturated designs. Stepwise often
fails to provide proper models, due to the multicollinearity existing in
supersaturated designs. A new proposed biased estimation technique for
analyzing the data of the supersaturated designs is proposed. The technique
combines the ridge regression estimation with a modified best subset variable
selection procedure, to select the significant factors in the model. Designs
of experiments are developed for different configurations of factor settings
in the true model. The performance measures are the observed Type I and
Type II average errors. The designed of experiments are applied to the
both the stepwise regression and the proposed method. Analysis of designed
experiments study is conducted on the different factors affecting Type
I and Type II errors. A comparison of results shows that the proposed method
performs better than the stepwise methods.