Archive for ABC

Bayesian optimization for likelihood-free inference of simulator-based statistical models [guest post]

Posted in Books, Statistics, University life with tags , , , , , , , on February 17, 2015 by xi'an

[The following comments are from Dennis Prangle, about the second half of the paper by Gutmann and Corander I commented last week.]

Here are some comments on the paper of Gutmann and Corander. My brief skim read through this concentrated on the second half of the paper, the applied methodology. So my comments should be quite complementary to Christian’s on the theoretical part!

ABC algorithms generally follow the template of proposing parameter values, simulating datasets and accepting/rejecting/weighting the results based on similarity to the observations. The output is a Monte Carlo sample from a target distribution, an approximation to the posterior. The most naive proposal distribution for the parameters is simply the prior, but this is inefficient if the prior is highly diffuse compared to the posterior. MCMC and SMC methods can be used to provide better proposal distributions. Nevertheless they often still seem quite inefficient, requiring repeated simulations in parts of parameter space which have already been well explored.

The strategy of this paper is to instead attempt to fit a non-parametric model to the target distribution (or in fact to a slight variation of it). Hopefully this will require many fewer simulations. This approach is quite similar to Richard Wilkinson’s recent paper. Richard fitted a Gaussian process to the ABC analogue of the log-likelihood. Gutmann and Corander introduce two main novelties:

  1. They model the expected discrepancy (i.e. distance) Δθ between the simulated and observed summary statistics. This is then transformed to estimate the likelihood. This is in contrast to Richard who transformed the discrepancy before modelling. This is the standard ABC approach of weighting the discrepancy depending on how close to 0 it is. The drawback of the latter approach is it requires picking a tuning parameter (the ABC acceptance threshold or bandwidth) in advance of the algorithm. The new approach still requires a tuning parameter but its choice can be delayed until the transformation is performed.
  2. They generate the θ values on-line using “Bayesian optimisation”. The idea is to pick θ to concentrate on the region near the minimum of the objective function, and also to reduce uncertainty in the Gaussian process. Thus well explored regions can usually be neglected. This is in contrast to Richard who chose θs using space filling design prior to performing any simulations.

I didn’t read the paper’s theory closely enough to decide whether (1) is a good idea. Certainly the results for the paper’s examples look convincing. Also, one issue with Richard‘s approach was that because the log-likelihood varied over such a wide variety of magnitudes, he needed to fit several “waves” of GPs. It would be nice to know if the approach of modelling the discrepancy has removed this problem, or if a single GP is still sometimes an insufficiently flexible model.

Novelty (2) is a very nice and natural approach to take here. I did wonder why the particular criterion in Equation (45) was used to decide on the next θ. Does this correspond to optimising some information theoretic quantity? Other practical questions were whether it’s possible to parallelise the method (I seem to remember talking to Michael Gutmann about this at NIPS but can’t remember his answer!), and how well the approach scales up with the dimension of the parameters.

Inference for stochastic simulation models by ABC

Posted in Books, Statistics, University life with tags , , , , , on February 13, 2015 by xi'an

Hartig et al. published a while ago (2011) a paper  in Ecology Letters entitled “Statistical inference for stochastic simulation models – theory and application”, which is mostly about ABC. (Florian Hartig pointed out the paper to me in a recent blog comment. about my discussion of the early parts of Guttman and Corander’s paper.) The paper is largely a tutorial and it reminds the reader about related methods like indirect inference and methods of moments. The authors also insist on presenting ABC as a particular case of likelihood approximation, whether non-parametric or parametric. Making connections with pseudo-likelihood and pseudo-marginal approaches. And including a discussion of the possible misfit of the assumed model, handled by an external error model. And also introducing the notion of informal likelihood (which could have been nicely linked with empirical likelihood). A last class of approximations presented therein is called rejection filters and reminds me very much of Ollie Ratman’s papers.

“Our general aim is to find sufficient statistics that are as close to minimal sufficiency as possible.” (p.819)

As in other ABC papers, and as often reported on this blog, I find the stress on sufficiency a wee bit too heavy as those models calling for approximation almost invariably do not allow for any form of useful sufficiency. Hence the mathematical statistics notion of sufficiency is mostly useless in such settings.

“A basic requirement is that the expectation value of the point-wise approximation of p(Sobs|φ) must be unbiased” (p.823)

As stated above the paper is mostly in tutorial mode, for instance explaining what MCMC and SMC methods are. As illustrated by the above figure. There is however a final and interesting discussion section on the impact of estimating the likelihood function at different values of the parameter. However, the authors seem to focus solely on pseudo-marginal results to validate this approximation, hence on unbiasedness, which does not work for most ABC approaches that I know. And for the approximations listed in the survey. Actually, it would be quite beneficial to devise a cheap tool to assess the bias or extra-variation due to the use of approximative techniques like ABC… A sort of 21st Century bootstrap?!

comments on reflections

Posted in pictures, Statistics, University life with tags , , , , , , on February 9, 2015 by xi'an

La Défense and Maison-Lafitte from my office, Université Paris-Dauphine, Nov. 05, 2011I just arXived my comments about A. Ronald Gallant’s “Reflections on the Probability Space Induced by Moment Conditions with Implications for Bayesian Inference”, capitalising on the three posts I wrote around the discussion talk I gave at the 6th French Econometrics conference last year. Nothing new there, except that I may get a response from Ron Gallant as this is submitted as a discussion of his related paper in Journal of Financial Econometrics. While my conclusion is rather negative, I find the issue of setting prior and model based on a limited amount of information of much interest, with obvious links with ABC, empirical likelihood and other approximation methods.

Bayesian computation: fore and aft

Posted in Books, Statistics, University life with tags , , , , , , , , , , , , on February 6, 2015 by xi'an

BagneuxWith my friends Peter Green (Bristol), Krzysztof Łatuszyński (Warwick) and Marcello Pereyra (Bristol), we just arXived the first version of “Bayesian computation: a perspective on the current state, and sampling backwards and forwards”, which first title was the title of this post. This is a survey of our own perspective on Bayesian computation, from what occurred in the last 25 years [a  lot!] to what could occur in the near future [a lot as well!]. Submitted to Statistics and Computing towards the special 25th anniversary issue, as announced in an earlier post.. Pulling strength and breadth from each other’s opinion, we have certainly attained more than the sum of our initial respective contributions, but we are welcoming comments about bits and pieces of importance that we miss and even more about promising new directions that are not posted in this survey. (A warning that is should go with most of my surveys is that my input in this paper will not differ by a large margin from ideas expressed here or in previous surveys.)

Bayesian optimization for likelihood-free inference of simulator-based statistical models

Posted in Books, Statistics, University life with tags , , , , on January 29, 2015 by xi'an

Michael Gutmann and Jukka Corander arXived this paper two weeks ago. I read part of it (mostly the extended introduction part) on the flight from Edinburgh to Birmingham this morning. I find the reflection it contains on the nature of the ABC approximation quite deep and thought-provoking.  Indeed, the major theme of the paper is to visualise ABC (which is admittedly shorter than “likelihood-free inference of simulator-based statistical models”!) as a regular computational method based on an approximation of the likelihood function at the observed value, yobs. This includes for example Simon Wood’s synthetic likelihood (who incidentally gave a talk on his method while I was in Oxford). As well as non-parametric versions. In both cases, the approximations are based on repeated simulations of pseudo-datasets for a given value of the parameter θ, either to produce an estimation of the mean and covariance of the sampling model as a function of θ or to construct genuine estimates of the likelihood function. As assumed by the authors, this calls for a small dimension θ. This approach actually allows for the inclusion of the synthetic approach as a lower bound on a non-parametric version.

In the case of Wood’s synthetic likelihood, two questions came to me:

  • the estimation of the mean and covariance functions is usually not smooth because new simulations are required for each new value of θ. I wonder how frequent is the case where we can always use the same basic random variates for all values of θ. Because it would then give a smooth version of the above. In the other cases, provided the dimension is manageable, a Gaussian process could be first fitted before using the approximation. Or any other form of regularization.
  • no mention is made [in the current paper] of the impact of the parametrization of the summary statistics. Once again, a Cox transform could be applied to each component of the summary for a better proximity of/to the normal distribution.

When reading about a non-parametric approximation to the likelihood (based on the summaries), the questions I scribbled on the paper were:

  • estimating a complete density when using this estimate at the single point yobs could possibly be superseded by a more efficient approach.
  • the authors study a kernel that is a function of the difference or distance between the summaries and which is maximal at zero. This is indeed rather frequent in the ABC literature, but does it impact the convergence properties of the kernel estimator?
  • the estimation of the tolerance, which happens to be a bandwidth in that case, does not appear to be processed in this paper, which could explain for very low probabilities of acceptance mentioned in the paper.
  • I am lost as to why lower bounds on likelihoods are relevant here. Unless this is intended for ABC maximum likelihood estimation.

Guttmann and Corander also comment on the first point, through the cost of producing a likelihood estimator. They therefore suggest to resort to regression and to avoid regions of low estimated likelihood. And rely on Bayesian optimisation. (Hopefully to be commented later.)

brief stop in Edinburgh

Posted in Mountains, pictures, Statistics, Travel, University life, Wines with tags , , , , , , , , on January 24, 2015 by xi'an

Edinburgh1Yesterday, I was all too briefly in Edinburgh for a few hours, to give a seminar in the School of Mathematics, on the random forests approach to ABC model choice (that was earlier rejected). (The slides are almost surely identical to those used at the NIPS workshop.) One interesting question at the end of the talk was on the potential bias in the posterior predictive expected loss, bias against some model from the collection of models being evaluated for selection. In the sense that the array of summaries used by the random forest could fail to capture features of a particular model and hence discriminate against it. While this is correct, there is no fundamental difference with implementing a posterior probability based on the same summaries. And the posterior predictive expected loss offers the advantage of testing, that is, for representative simulations from each model, of returning the corresponding model prediction error to highlight poor performances on some models. A further discussion over tea led me to ponder whether or not we could expand the use of random forests to Bayesian quantile regression. However, this would imply a monotonicity structure on a collection of random forests, which sounds daunting…

My stay in Edinburgh was quite brief as I drove to the Highlands after the seminar, heading to Fort William, Although the weather was rather ghastly, the traffic was fairly light and I managed to get there unscathed, without hitting any of the deer of Rannoch Mor (saw one dead by the side of the road though…) or the snow banks of the narrow roads along Loch Lubnaig. And, as usual, it still was a pleasant feeling to drive through those places associated with climbs and hikes, Crianlarich, Tyndrum, Bridge of Orchy, and Glencoe. And to get in town early enough to enjoy a quick dinner at The Grog & Gruel, reflecting I must have had half a dozen dinners there with friends (or not) over the years. And drinking a great heather ale to them!

not Bayesian enough?!

Posted in Books, Statistics, University life with tags , , , , , , , on January 23, 2015 by xi'an

Elm tree in the park, Parc de Sceaux, Nov. 22, 2011Our random forest paper was alas rejected last week. Alas because I think the approach is a significant advance in ABC methodology when implemented for model choice, avoiding the delicate selection of summary statistics and the report of shaky posterior probability approximation. Alas also because the referees somewhat missed the point, apparently perceiving random forests as a way to project a large collection of summary statistics on a limited dimensional vector as in the Read Paper of Paul Fearnhead and Dennis Prarngle, while the central point in using random forests is the avoidance of a selection or projection of summary statistics.  They also dismissed ou approach based on the argument that the reduction in error rate brought by random forests over LDA or standard (k-nn) ABC is “marginal”, which indicates a degree of misunderstanding of what the classification error stand for in machine learning: the maximum possible gain in supervised learning with a large number of classes cannot be brought arbitrarily close to zero. Last but not least, the referees did not appreciate why we mostly cannot trust posterior probabilities produced by ABC model choice and hence why the posterior error loss is a valuable and almost inevitable machine learning alternative, dismissing the posterior expected loss as being not Bayesian enough (or at all), for “averaging over hypothetical datasets” (which is a replicate of Jeffreys‘ famous criticism of p-values)! Certainly a first time for me to be rejected based on this argument!


Get every new post delivered to your Inbox.

Join 772 other followers