The trip to Hamburg had started inauspiciously: a heart attack (someone else’s heart) in the métro (RER) has frozen the train traffic completely on Tuesday and I was lucky to find a taxi that managed to drive me to the airport in the nick of time. As there were warnings of strike in the Hamburg airport today, I decided to pack early and left DESY long enough in advance to reach the aiport by public transportation: it is only once I cleared security and sat at the gate that I realised I had forgotten my PC power box/cord in my room at DESY. What a drag! Anyway, I managed to buy a universal adapter in Paris on my way back from the airport and still to attend Adam Johansen’s seminar on Rao-Blackwellisation of particle filters at the Big’MC seminar. An interesting exploitation of missing variable structures within missing variables in a hidden Markov chain! (I missed the reason for the O(NM) computing time, though.) On the way home, I reflected on how little I had seen of Hamburg: a nice train system, a green and pleasant suburban area south of DESY towards the Elben (Groß Flottbeck), while running this morning, and…the airport! I wish I had had the opportunity and the time to get a glimpse of downtown Hamburg.
Archive for Rao-Blackwellisation
This is the last series of lectures for my AMSI Lecture tour: in Brisbane, I gave the Rao-Blackwellisation talk this afternoon at the University of Queensland and will give tomorrow both the public lecture (for the second time) and the ABC for model choice at QUT. This will then see the end to this (almost) exhaustive if exhausting tour of Eastern Australian university towns… Brisbane has this great feature of connecting QUT with UQ by a fast boat, which meant I could work with Kerrie Mengersen on the revision of our ABCel paper in the morning and be at UQ ½ an hour later. A good thing as there are only three days left till I leave Australia after this seven week visit… (I am afraid there is no free time left for sampling the Brisbane dumplings!)
After a short two-day visit to Canberra, I am now heading to Brisbane, last stop of my Australian Tour… I gave my Kyoto talk at ANU, failing to convince my guest Alan Welsh that ABC was a completely justified Bayesian inference method (with the sound argument that it failed the conditionality principle), had a great run in the morning to the top of Black Mountain, witnessing a very nice inversion shown below,
ate a great Indian dinner, and was driven by Alan the next morning to Goulburn for a joint meeting (“Goulburn 19″) between the statisticians of Canberra and of the University of Wollongong. This was a fairly interesting meeting, taking place in a motel conference room. Luxury, luxury!, the room even had a log fire (with the foreseeable consequence of inducing one of my customary and embarrassing dozes!) Actually, this seemed to be a signature of the place as the terrific Trapper’s Bakery across the street also had a fireplace… This meeting was very informal, friendly, and interactive, making me regret I had given there my most technical talk on Rao-Blackwellisation instead of one on ABC model choice.
As I left Banff in the early afternoon of Thursday, I only attended the morning session of the BIRS 12w5105 workshop (just like yesterday since there was a free afternoon!). Yesterday morning, Gareth Roberts gave a brilliant talk on importance sampling for diffusions, explaining why he could use importance functions that were not absolutely continuous wrt the target distribution. I also like Jun Liu’s talk on sequential Monte Carlo. This idea of using partial or controlled rejection sampling is quite interesting indeed, even though I always wonder at why rejection is at all necessary since it introduces extra (white?) noise… Rao-Blackwellised solutions, as mentioned by Jun, should be more efficient. An issue certainly not fit for an airport lounge! This morning, the first talk by Hélène Massam was also close to my interests, since it dealt simultaneously with model choice and Bayes factors and their limit as the prior becomes improper and graphical models… I should read the recent work of Hélène on this. Her talk also mentioned the issue of computing simultaneously thousands of Bayes factors, a good challenge (even though ABC seems ill-fitted to the task). Jim Hobert gave the last talk of the morning on geometric convergence and its verification by construction a potential function: as usual, Jim’s slides were fantastically pedagogical, make sure to watch them! I am very much sorry to have missed the three afternoon talks, but will watch them over the weekend, thanks to this new incredible feature of BIRS where all videos are now on-line.
Overall, this week at BIRS, staying at the Banff Centre, and doing mountaineering, has been exhilarating if exhausting! The incredible conditions offered by the Centre go beyond the reasonable and it takes alas little time to get used to them, from the sport facilities, with my very early morning swim (6am), to the superb Sally Borden cafetaria, to the auditorium in the TransCanada Pipeline Pavilion with its video-recording facilities, and I am only sorry I could not find a climbing partner to enjoy the indoor climbing wall! Scientifically speaking, the structure of the meeting allowed for intense exchanges (even though I did not exploit those to their full extent, due to my early bedtime, trying to keep on French time, a habit which allowed me to review Pierre Simon Laplace’s book, Théorie Analytique des Probabilités, over the week, during the wee hours of the morning…, but also led to almost sure micro-naps during the talks…!)
Although this was only a half-day of talks, the third day of the workshop was equally thought-challenging and diverse. (I managed to miss the ten first minutes by taking a Line 3 train to 125th street, having overlooked the earlier split from Line 1… Crossing south Harlem on a Sunday morning is a fairly mild experience though.) Jean-Marc Azaïs gave a personal recollection on the work of Mario Wschebor, who passed away a week ago and should have attended the workshop. Nan Chen talked about the Monte Carlo approximation of quantities of the form
which is a problem when f is non linear. This reminded me (and others) of the Bernoulli factory and of the similar trick we use in the vanilla Rao-Blackwellisation paper with Randal Douc. However, the approach was different in that the authors relied on a nested simulation algorithm that did not adapt against f. And did not account for the shape of f. Peter Glynn, while also involved in the above, delivered a talk on the initial transient that showed possibilities for MCMC convergence assessment (even though this is a much less active area than earlier). And, as a fitting conclusion, the conference organiser, Jingchen Liu gave a talk on non-compatible conditionals he and Andrew are using to handle massively-missing datasets. It reminded me of Hobert and Casella (1996, JASA) of course and also of discussions we had in Paris with Andrew and Nicolas. Looking forward to the paper (as I have missed some points about the difference between true and operational models)! Overall, this was thus a terrific workshop (I just wish I could have been able to sleep one hour more each night to be more alert during all talks!) and a fantastic if intense schedule fitting the start of the semester and of teaching (even though Robin had to teach my first R class in English on Friday). I also discovered that several of the participants were attending the Winter Simulation Conference later this year, hence another opportunity to discuss simulation strategies together.