19 T.W. Alexander Drive P.O.Box 14006 Research Triangle Park, NC 27709-4006 Tel: 919.685.9350 Fax: 919.685.9360 info@samsi.info

Population Monte Carlo Working Group - Meeting Summary Page

July
June
May
November
October
September

July 2008

Remote attendees: Bob, Ioanna, Jessi, Julien, Sylvain

Summary of ABC in Paris workshop
Summary by Julien. My apologies to those of the speakers that I did not pay enough attention to ! (I invoke marathon schedule, Paris June heat in the amphitheater, and an still-partial recovery Ph.D. defense the day before, as weak excuses).

• Pretty intense day, with a lot of great speakers and auditors (60 persons according to the organizer).
• The slides of the talks are available online, as well as some videos.
• Note also that Christian Robert preliminarily discussed several of the presented articles.

• Ajay Jasra on adaptive SMC samplers: key trick is choosing the kernels so that the resampling weights simplify and depend only on the ancestor particles
• Tina Toni: Introductory talk on ABC for Dynamical systems, with special interest on model selection
• Mark Briers: A rejoinder on the very "intense" topic of the biased weights, to make everybody "one happy ABC family again" (sic). Great job, not an easy task ! Note that he mentioned transdimensional moves in SMC samplers, this will be relevant to our subgroup Model Selection and Transdimensional Inference.
• Christoph Leuneberger: most relevant to our subgroup Applications to Population Genetics.
• Oliver Ratmann: talk about his article recently accepted in PNAS, which uses the distance between the summary statistics as a tool for model criticism. Discussed by Christian Robert.
• Jean-Michel Marin: joint work with Aude from our subgroup on Population Genetics, about model selection for Gibbs Random Fields.
• David Badding's, Paul Fearnhead's and Olivier François' respective talks: all focus on the design of the summary statistics, especially through regression. Paul Fearnhead especially suggests to use linear regression on powers of the observation, with rationale rooted in Taylor expensions.
• Marc Beaumont: application to population genetics.
• Arnaud Estoup: presentation of a user-friendly software, especially aimed at population biologists, with a very usable graphical interface.

Status report from all sub groups

• ABC: Julien at last writing a note on his ideas on Auxiliary SMC samplers. Seem to indicate that backward kernel can transport from a continuous space to a discrete space, even when the state-space is continuous (to be taken with great care, this is Work In Progress).
• SMC samplers for bayesian analysis of massive dataset: No progress this week, Julien's, Ioanna's, and Artin's schedules are very conflicting in these conferences' times !
• Applications to population genetics and epidemiology: Regarding the application to epidemiology, Bob and Aude are currently focusing on the way to propose ancestors. Aude had had prepared some slides but could not join the meeting due to network problems.
• Adaptive SMC samplers: On hold until Arnaud and Julien can meet. Note that most of the work in the ABC subgroup will serve as a starting point.
• Model selection and transdimensional inference: Julien put Mark Briers and Giovanni in touch about the transdimensional moves mentioned in the summary above. Jan will be back around July 20 (cruising the world for conferences). When he will be back, Jessi will present on her use of SMC for fiducial inference.
• Theoretical review: Anand is going on steadily on the theoretical articles mentioned last meeting. Julien and Eric Moulines will be visiting him on July 16-17. During ABC in Paris, informal discussions indicated that a few people are currently working on theoretical proofs of convergence for ABC-SMC.

Next meeting Thursday July 23rd, 1pm UTC-0400.
Because in two weeks a conference will take place in Cornell that prevents Julien from joining, we will meet in three weeks, on July 23rd. Until then:

• Julien will see if Giovanni and Mark succeeded in getting in touch about transdimensional moves,
• Aude will send her slides, so that we can all read them,
• Julien and Bob will see if David is still willing to get onboard the epidemiology subgroup.

June 2008

Local attendees: Ioanna
Remote attendees: Bob, Jessi, Julien

Status report from all sub groups
Each subgroup leader quickly presented the updates of the subgroup (in order of presentation):

• ABC: see Julien's slides, slide 3. Working note from Oliver Ratmann also shown.
• SMC samplers for bayesian analysis of massive dataset: No progress this week.
• Applications to population genetics: Bob, Aude and David are now in contact, with very good advances on the bibliography and the problem statement.
• Adaptive SMC samplers: No progress this week, will most likely be the case until Arnaud and Julien can meet.
• Model selection and transdimensional inference: see Jessi's infos in the slides, slide 4. She will be presenting at next meeting her work with Jan on simulation methods for parameter estimation fiducial inference, with an added twist that has made resampling impossible so far.
• Theoretical review: No report this week.

Next meeting Thurday July 2nd, 1pm UTC-0400.

Local attendees: Julien, Ioanna, Sourish
Remote attendees: Anand, Bob, Jessi

Status report from all sub groups
Each subgroup leader presented the updates of the subgroup (in order of presentation):

• ABC: Julien and Olli are currently getting on grips with the several ABC-SMC existing variants in the abundant and very recent literature, to have a sound basis.
• SMC samplers for bayesian analysis of massive dataset: Recoding the algorithms in a more powerful language (Matlab) than R increased computational efficiency by an order of magnitude. The challenge hences will come more from the high-dimensionality of the data that its (finally not so) massive nature.
• Applications to population genetics: Aude has been doing a lot of reading on the topic. Bob thinks of linking the SMC approaches used in Population Genetics to some Spatial Epidemiologic models, as they share common properties of a known forward dynamic and a need for reconstructing the unknown backward one. Sourish mentions David experience and earlier work in epidemiology, so it may be very relevant to have join the subgroup. Besides, this could have deep connections with next year's SAMSI program.
• Adaptive SMC samplers: In standby until Arnaud and Julien meet to work on it. Recent litterature show it is definitely in the air, we shall not wait too long !
• Model selection and transdimensional inference: Neither Jan nor Giovanni could make it today, so no status update.
• Theoretical review: Anand could not get in touch with Jan, but is progressing on the review of asymptotic results for SMC, with a grad student of his.

Next meeting Thurday June 18th, 1pm UTC-0400.

May 2008

Local attendees: Julien, David, Jan, Ioanna, Sourish
Remote attendees: Anand, Arnaud, Robert, Sylvain
Giovanni, Aude, Gareth, Mark, Jessi, warned they could not make it at this time today.

Presentation of the sub groups
Each subgroup leader presented the aims of the subgroup (in order of presentation):

• Applications to population genetics: Bob's slides.
• SMC samplers for bayesian analysis of massive dataset: Julien's slides
• Adaptive SMC samplers: Julien's slides
• ABC: Julien was out of slides, and hence summarized that the point of this subgroup is, in a first time at least, to put together the results from his Ph.D. thesis with those of Oliver Ratmann's Ph.D. thesis.
• Model selection and transdimensional inference: Jan made some oral explanations with the help of the board.
• Theoretical review: Anand's slides. Discussion followed on the articles he cited.

Organization from now on
The subgroups will now work autonomously, with regular common status updates.
The subgroup leaders will contact the members of the groups as signed on in the updated list. Please feel free to contact them too to state your interest !
Two points about schedule:

1. We will meet every two weeks for short meetings (no more than 30 minutes) to have a global status update. It would be really helpful if each subgroup could ensure that at least one member will (even remotely) assist. For now on, we will keep this meeting day/time, every other Thursday at 1pm UTC-0400.
2. We aim at results (from short preliminary technical reports to preprints) before SAMSI's annual report which is roughly mid-September. Of course, a SAMSI program never stops but evolves to independent interactions, so the work started now can continue way longer. However, it is good to have this deadline in mind...

Next meeting Thurday May 21st, 1pm UTC-0400.

November 2008

Locally present: Arnaud, Geneviève (Lefebvre), Jan, Julien, Mark (Huber)
Remote: Giovanni, Sylvain

Geneviève Lefebvre's lecture

Genevi`ve Lefebvre, from from Université du Québec à Montréal, lectured on one of the hot topic of your working group, namely designing the cooling schedule in SMC Sampler, based on Kullback-Leibler and J-divergence.

• A divergence approach for the tempering schedule in Sequential Monte Carlo Sampling
• Slides are available.

This works is part of her PhD thesis [16] Lefebvre (2007).

Organizational matters

No lecture on next week, due to Thanksgiving holiday.

October 2008

Locally present: Arnaud, Christian (Robert), Jessie, Roman, Thomas (Flury)
Remote: Sylvain
Julien warned in advance that he could not join

Jan Hannig's lecture

• On Generalized Fiducial Inference
• Slides, board

Locally present: Arnaud, Gareth, Jan, Julien, Luke, Nathan, Roman
Remote: Anand, Ernest, Sylvain

Arnaud Doucet's lecture on SMC samplers

Arnaud Doucet gave a lecture on Sequential Monte Carlo Samplers, whose slides are available. It relies on his article [1] Del Moral, Doucet, Jasra (2006), with additional material.

Challenge to focus on

At the end of this lecture, Arnaud pointed to a challenge which could form the focus of a subgroup, on clever design of the cooling schedule and the backward kernels involved in the algorithm, and on ways to assess their efficiency, going further than the currently used Effective Sample Size. He will soon make available a few pages summarizing this.

For next meeting

Jan Hannig will be lecturing next week. If you are willing to lecture on the week after, please contact Arnaud or Julien.

Locally present: Artin, Gareth, Julien, Nathan, Thomas
Remote: Mark B., Sylvain

Julien Cornebise's lecture on quality criterions and adaptive SMC

Julien lectured on Analysis of quality criterions and adaptive methods in sequential importance sampling.
The slides are available, along with the article [13] Cornebise, Moulines, Olsson (2008) on which this talk is based. Julien also added the reference [14] Douc, Moulines (2008) for useful convergence results (with weaker assumptions than those commonly found in the litterature) based on a step-by-step analysis of SMC algorithms, and [15] Douc, Moulines, Olsson (2008) for an application of these results to auxiliary particle filter, also used in the article [13].

For next meeting

Though this is unfortunately becoming as running gag, we are looking for volunteers for lecturing. Please contact Julien Cornebise and Arnaud Doucet if you would like to talk.

Locally present: Arnaud, Gareth, Hedibert, Jan, Jaya, Julien, Nathan, Roman, Sylvain, Thomas
Remote: Anand

Sylvain Rubenthaler's lecture on propagation of chaos

Sylvain lectured on Tree based functional expansions for particle models, a much extended version of the talk he gave in Duke last week (now including the sktech of the proof and the applications).
His slides are available (first part of the talk) along with the further developments on the whiteboard (second part of the talk). This lecture relies on his article [12] Del Moral, Patras, Rubenthaler (2006), soon to be published.

For next meeting

Again, looking for a volunteer for lecturing. We have several lecturers for November, but none for the end of October, so please contact Julien Cornebise and Arnaud Doucet if you would like to talk.

Locally present: Artin, Ioanna, Jan, Jonathan, Julien
Remote: Anand, Bacun, Bob, Bingxin, Cajo (connexion problem), Ernest, Mark B., Steve
Arnaud will be back next week.

Bob Griffith's lecture on population genetics

Bob lectured on Ancestral inference on coalescent histories, which is one of the applications of interest of our working group. His slides are available for download, and include a rich list of references.

When asked for a short list of references, Bob advises:

• [9] Griffiths (2002) for a review of the topic,
• [6] Griffiths and Tavaré (1994), [8] Stephens and Donnelly (2000), [10] De Iorio and Griffiths (2004), and [11] De Iorio et al. (2005) as articles on the topic, and
• [7] Harding et al. (1997) for an application on real dataset.

For next meeting

September 2008

Locally present: Sylvain, Nathan, Gareth, Julien, Ola, Mark H., + 1.
Remote: Anand, Baqun, Mark B., Steve.
Arnaud was on his way to Japan, so could not join us.

Sylvain Rubenthaler's lecture on Boltzmann particles

Sylvain lectured on the whiteboard pp. 98-106 of [3], about Particle approximations for Boltzmann equation (mollified). The whole lecture is available for download.

For next meetings

First and foremost, we are looking for a lecturer for next week, who could present an informal introduction to an article or a field of interest to the working group: any volunteer, please contact Arnaud and Julien. [Update 9/26 1:20PM] Robert Griffiths will give an introductory lecture to SMC in Population Genetics, whose title will be Sequential Importance Sampling in Population Genetics: Looking backwards in time.

For potential applications of SMC as alternative to MCMC, Cajo (who could not make it today) points to

• [4] Bink et al. (2008) as an introduction to QTL analysis, and
• [5] Jain et al. (2007) about haplotype reconstruction.

We will talk about these in a future meeting. As usual, next meeting is same time, same day, next week. Speaker will be announced beforehand.

Introductory lecture by Arnaud Doucet on SMC: A Crash Course on SMC Methods (slides), a one-hour-and-a-half version of his eight-hours grad course. It only very briefly mentionned Population Monte-Carlo methods in themselves, and is more of a review of the basis so that we all share a common basis.
Sylvain Rubenthaler then gave the references [2] Graham and Meleard (1997) and [3] Graham and Meleard (1999), that he will present in a talk at the next meeting. The 1999 article are lecture notes from a summer school, based on the 1997 article: it is hence more lengthy and more detailed, but shares the same core content.

The minutes of the meeting are available in PDF.

Working Group Home Page

Entire site © 2002-2008, Statistical and Applied Mathematical Sciences Institute. All Rights Reserved.

Last modified