19 T.W. Alexander Drive
P.O.Box 14006 Research Triangle Park, NC 277094006 Tel: 919.685.9350 Fax: 919.685.9360 info@samsi.info 
Methodology
A
Working Group in Development, Assessment and Utilization of Complex
Computer Models
Group Leaders: Susie Bayarri ( Univ. of Valencia) and Robert Wolpert (Duke)
Password Protected Site (future)
Announcements 
Suggested
Readings 
Meeting
Activities 
Group Members 
Meeting Time: Monday 11:00 
1:00 Remote access for the meeting is available. WebEx and teleconference dialin instructions here. Via teleconference at 9196859338. Note: Webcam access is not available
this year. 
Treed
GPs and some of the adaptive design computations are in the R
package tgp, which can be accessed from within R using
install.packages("tgp") or from the CRAN
website. 7. Screening. This is an area that has not appeared much in our discussions. We discussed it some in the October joint Methodology and Engineering Workshop. The paper : Variable Selection for Gaussian Process Models in Computer Experiments (by Crystal Linkletter, Derek Bingham, Nicholas Hengartner, David Higdon, and Kenny Q. Ye) describes the Bayesian approach that Derek Bingham was referring to. 8. Terminology. There seems to be terms like "validation" that are used with different meanings by different groups. Tony O'Hagan gave a summary of his proposed terminology in the final part of his 1172006 talk. Tom Loredo told us that actually there is a specific DOE whitepaper defining how DOE expects its contractors to use the terms verification, validation, and uncertainty quantification. It is available directly from the DOE at this URL. This document was prepared for the DOE's Predictive Science Academic Alliance Program (PSAAP), a major initiative "to establish validated, largescale, multidisciplinary, simulationbased 'Predictive Science' as a major academic and applied research program." You can learn more about PSAAP from its URL. 9. Quantile Regression. A
couple summary
papers on quantile regreesions are those of Koenker
and Hallock (2001) and Cade
and Noon (2003); an implementation in R appears in Koenker
(2006) . 10. NonGaussian Spatial
Processes. Diggle et al.
(1998) dicusses Generalized linear mixed model for Geostatistics. 11. Tapering. Kaufman (2007) is considering
using Covariance Tapering method to build efficient emulators for
computer experiments. Two revelant papers are: Wendland (1995) introduces
Piecewise polynomial, positive definite and compactly supported radial
functions of minimal degree. Wendland
(1998) discusses error estimates for interpolation by compactly
supported radial basis functions of minimal degree. 12. The link to the Gaussian
Process website is here.
13. Use of derivatives
(list provided by Tom Loredo).
1  O'Hagan, A. (1992). Some
Bayesian numerical analysis (with discussion). In Bayesian
Statistics 4, (J. M. Bernardo et al Eds.), 345363. Oxford University
Press.

September
2006 
October
2006 
November
2006 
December
2006 
January 2007  Febuary 2007  March 2007  April 2007 
Date  Topics & Readings  Note 
4012007 
We will be having the joint SAMSI/MUCM workshop, and hence there will be no meeting of the working group. Please be reminded that there will be webex access (different from the usual one) for those of you who have email Margaret about it.  
4162007 
We'll have an interesting
brainstorming session with Michael Goldstein about the informal theme
"What I'm finding currently interesting and challenging about computer
model problems". This promises to be a very stimulating
discussions, and I hope you can make it. There will probably be no
slides, but we'll activate the Smart Board in case we do some writing. 

4232007 
We
will have another discussion session with Michael Golstein. This time
he will be addressing the issue of relating coarse computer models with
fine ones, and relating a simulator to reality. The ideas develop from
the work in the Linear
Bayes Methodology reading list. Several papers there discuss
issues of linking up simulators at different resolutions. Perhaps the
last paper on the list
"Constructing partial prior
specifications for models of complex physical systems." might be helpful, because it doesn't get involved in trying to solve problems with the combined specification. On the other hand "Pressure matching for hydocarbon reservoirs: a case study in the use of Bayes linear strategies for large computer experiments" does describe (more or less) how we actually went about linking two specific models and our current paper "Reified Bayesian Modelling and Inference for Physical Systems" pushes these ideas a lot further. These are some slides on `reification'. 

4252007 
We will meet at 10:0011:30.
Professor Jim Zidek is going to give a talk as the following. Here are the slides. Title: Bayesian melding in the hunt for the elusive PRB level Speaker: Jim Zidek, U British Columbia Abstract: A PRB (Policy related background) level of a criterion pollutant such as ozone addressed in US regularity policy, is its (hypothetical) level if there were no anthropogenic sources in North America. It is a foundation above which to set regulatory strandards. However it cannot be measured since pristine areas no longer exist anywhere over the continental US. Hence it is has been inferred (predicted) by deterministic CTMs (chemical transport models). That strategy generates a lot of concerns, in particular about the accuracy and bias in the predictions. To address those concerns, (seemingly favorable) comparisons have made of those predictions and measurements with the anthrosources turned ON, to enhance the credibility of the predictions when those sources are turned OFF. The Bayesian melding approach of Fuentes and Raftery offers a refined approach for comparing CTM outputs and that approach will be the main topic of this presentation. I will describe its potential for use in the hunt for the PRB level and our experiences with the method, in work that stemmed from my 6 month visit to SAMSI in 2003. In particular it was stimulated by conversations with Duke's Prasad Kasibhatla, who supplied for the MAQSIP CTM outputs as well as AIRS hourly ozone concentrations we used. I will in particular describe some of the joy and DESPAIR of using MCMC for parameter high dimensional parameter vector. 

4302007 
No
meeting. 

512007 
We join the Engineering Methodology working group for a detailed exposition of a novel method to emulate dynamic models. Peter Reichert and Gentry White will be presenting.  
572007 
Please send email to the group leaders if you have new results and/or questions which would like to discuss with the group by Thursday 12:00 noon. Otherwise, we will only meet on Tuesday (12:30  14:00). Email notification will be sent if meeting will be held on Monday.  
582007 
We meet again jointly with the engineering methodology working group (on their usual time slot on Tuesday 12:30  14:00) to continue discussion of the dynamic emulator. This might be the last of our regular scheduled meetings, but if someone has a different possibility in mind, or interesting ideas for future activities, or whatever, please take them to the meeting on Tuesday. 