## Archive for adaptive MCMC methods

Posted in Books, Statistics, University life with tags , , , , , , , , , , on July 3, 2015 by xi'an

Heiko Strathmann, Dino Sejdinovic, Samuel Livingstone, Zoltán Szabó, and Arthur Gretton arXived paper about Kamiltonian MCMC generated comments from Michael Betancourt, Dan Simpson and myself, which themselves induced the following reply by Heiko, detailed enough to deserve a post of its own.

We certainly agree that the naive approach of using a non-parametric kernel density estimator on the chain history (as in [Christian’s book, Example 8.8]) as a *proposal* fails spectacularly on simple examples: the probability of proposing in unexplored regions is extremely small, independent of the current position of the MCMC trajectory. This is not what we do though. Instead, we use the gradient of a density estimator, and not the density itself, for our HMC proposal. Just like KAMH, KMC lite in fact falls back to Random Walk Metropolis in previously unexplored regions and therefore inherits geometric ergodicity properties. This in particular includes the ability to explore previously “unseen” regions, even if adaptation has stopped. I implemented a simple illustration and comparison here.

ABC example.
The main point of the ABC example, is that our method does not suffer from the additional bias from Gaussian synthetic likelihoods when being confronted with skewed models. But there is also a computational efficiency aspect. The scheme by Meeds et al. relies on finite differences and requires $2D$ simulations from the likelihood *every time* the gradient is evaluated (i.e. every leapfrog iteration) and H-ABC discards this valuable information subsequently. In contrast, KMC accumulates gradient information from simulations: it only requires to simulate from the likelihood *once* in the accept/reject step after the leapfrog integration (where gradients are available in closed form). The density is only updated then, and not during the leapfrog integration. Similar work on speeding up HMC via energy surrogates can be applied in the tall data scenario.

Approximating HMC when gradients aren’t available is in general a difficult problem. One approach (like surrogate models) may work well in some scenarios while a different approach (i.e. Monte Carlo) may work better in others, and the ABC example showcases such a case. We very much doubt that one size will fit all — but rather claim that it is of interest to find and document these scenarios.
Michael raised the concern that intractable gradients in the Pseudo-Marginal case can be avoided by running an MCMC chain on the joint space (e.g. $(f,\theta)$ for the GP classifier). To us, however, the situation is not that clear. In many cases, the correlations between variables can cause convergence problems (see e.g. here) for the MCMC and have to be addressed by de-correlation schemes (as here), or e.g. by incorporating geometric information, which also needs fixes as Michaels’s very own one. Which is the method of choice with a particular statistical problem at hand? Which method gives the smallest estimation error (if that is the goal?) for a given problem? Estimation error per time? A thorough comparison of these different classes of algorithms in terms of performance related to problem class would help here. Most papers (including ours) only show experiments favouring their own method.

GP estimator quality.
Finally, to address Michael’s point on the consistency of the GP estimator of the density gradient: this is discussed In the original paper on the infinite dimensional exponential family. As Michael points out, higher dimensional problems are unavoidably harder, however the specific details are rather involved. First, in terms of theory: both the well-specified case (when the natural parameter is in the RKHS, Section 4), and the ill-specified case (the natural parameter is in a “reasonable”, larger class of functions, Section 5), the estimate is consistent. Consistency is obtained in various metrics, including the L² error on gradients. The rates depend on how smooth the natural parameter is (and indeed a poor choice of hyper-parameter will mean slower convergence). The key point, in regards to Michael’s question, is that the smoothness requirement becomes more restrictive as the dimension increases: see Section 4.2, “range space assumption”.
Second, in terms of practice: we have found in experiments that the infinite dimensional exponential family does perform considerably better than a kernel density estimator when the dimension increases (Section 6). In other words, our density estimator can take advantage of smoothness properties of the “true” target density to get good convergence rates. As a practical strategy for hyper-parameter choice, we cross-validate, which works well empirically despite being distasteful to Bayesians. Experiments in the KMC paper also indicate that we can scale these estimators up to dimensions in the 100s on Laptop computers (unlike most other gradient estimation techniques in HMC, e.g. the ones in your HMC & sub-sampling note, or the finite differences in Meeds et al).

## Kamiltonian Monte Carlo [no typo]

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

Heiko Strathmann, Dino Sejdinovic, Samuel Livingstone, Zoltán Szabó, and Arthur Gretton arXived a paper last week about Kamiltonian MCMC, the K being related with RKHS. (RKHS as in another KAMH paper for adaptive Metropolis-Hastings by essentially the same authors, plus Maria Lomeli and Christophe Andrieu. And another paper by some of the authors on density estimation via infinite exponential family models.) The goal here is to bypass the computation of the derivatives in the moves of the Hamiltonian MCMC algorithm by using a kernel surrogate. While the genuine RKHS approach operates within an infinite exponential family model, two versions are proposed, KMC lite with an increasing sequence of RKHS subspaces, and KMC finite, with a finite dimensional space. In practice, this means using a leapfrog integrator with a different potential function, hence with a different dynamics.

The estimation of the infinite exponential family model is somewhat of an issue, as it is estimated from the past history of the Markov chain, simplified into a random subsample from this history [presumably without replacement, meaning the Markovian structure is lost on the subsample]. This is puzzling because there is dependence on the whole past, which cancels ergodicity guarantees… For instance, we gave an illustration in Introducing Monte Carlo Methods with R [Chapter 8] of the poor impact of approximating the target by non-parametric kernel estimates. I would thus lean towards the requirement of a secondary Markov chain to build this kernel estimate. The authors are obviously aware of this difficulty and advocate an attenuation scheme. There is also the issue of the cost of a kernel estimate, in O(n³) for a subsample of size n. If, instead, a fixed dimension m for the RKHS is selected, the cost is in O(tm²+m³), with the advantage of a feasible on-line update, making it an O(m³) cost in fine. But again the worry of using the whole past of the Markov chain to set its future path…

Among the experiments, a KMC for ABC that follows the recent proposal of Hamiltonian ABC by Meeds et al. The arguments  are interesting albeit sketchy: KMC-ABC does not require simulations at each leapfrog step, is it because the kernel approximation does not get updated at each step? Puzzling.

I also discussed the paper with Michael Betancourt (Warwick) and here his comments:

“I’m hesitant for the same reason I’ve been hesitant about algorithms like Bayesian quadrature and GP emulators in general. Outside of a few dimensions I’m not convinced that GP priors have enough regularization to really specify the interpolation between the available samples, so any algorithm that uses a single interpolation will be fundamentally limited (as I believe is born out in non-trivial scaling examples) and trying to marginalize over interpolations will be too awkward.

They’re really using kernel methods to model the target density which then gives the gradient analytically. RKHS/kernel methods/ Gaussian processes are all the same math — they’re putting prior measures over functions. My hesitancy is that these measures are at once more diffuse than people think (there are lots of functions satisfying a given smoothness criterion) and more rigid than people think (perturb any of the smoothness hyper-parameters and you get an entirely new space of functions).

When using these methods as an emulator you have to set the values of the hyper-parameters which locks in a very singular definition of smoothness and neglects all others. But even within this singular definition there are a huge number of possible functions. So when you only have a few points to constrain the emulation surface, how accurate can you expect the emulator to be between the points?

In most cases where the gradient is unavailable it’s either because (a) people are using decades-old Fortran black boxes that no one understands, in which case there are bigger problems than trying to improve statistical methods or (b) there’s a marginalization, in which case the gradients are given by integrals which can be approximated with more MCMC. Lots of options.”

## Bayesian computation: fore and aft

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

With 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.)

## self-healing umbrella sampling

Posted in Kids, pictures, Statistics, University life with tags , , , , , , , on November 5, 2014 by xi'an

Ten days ago, Gersende Fort, Benjamin Jourdain, Tony Lelièvre, and Gabriel Stoltz arXived a study about an adaptive umbrella sampler that can be re-interpreted as a Wang-Landau algorithm, if not the most efficient version of the latter. This reminded me very much of the workshop we had all together in Edinburgh last June. And even more of the focus of the molecular dynamics talks in this same ICMS workshop about accelerating the MCMC exploration of multimodal targets. The self-healing aspect of the sampler is to adapt to the multimodal structure thanks to a partition that defines a biased sampling scheme spending time in each set of the partition in a frequency proportional to weights. While the optimal weights are the weights of the sets against the target distribution (are they truly optimal?! I would have thought lifting low density regions, i.e., marshes, could improve the mixing of the chain for a given proposal), those are unknown and they need to be estimated by an adaptive scheme that makes staying in a given set the less desirable the more one has visited it. By increasing the inverse weight of a given set by a factor each time it is visited. Which sounds indeed like Wang-Landau. The plus side of the self-healing umbrella sampler is that it only depends on a scale γ (and on the partition). Besides converging to the right weights of course. The downside is that it does not reach the most efficient convergence, since the adaptivity weight decreases in 1/n rather than 1/√n.

Note that the paper contains a massive experimental side where the authors checked the impact of various parameters by Monte Carlo studies of estimators involving more than a billion iterations. Apparently repeated a large number of times.

The next step in adaptivity should be about the adaptive determination of the partition, hoping for a robustness against the dimension of the space. Which may be unreachable if I judge by the apparent deceleration of the method when the number of terms in the partition increases.

## MCqMC 2014 [day #3]

Posted in pictures, Running, Statistics, Travel, University life, Wines with tags , , , , , , , , , , , , , , , , on April 10, 2014 by xi'an

As the second day at MCqMC 2014, was mostly on multi-level Monte Carlo and quasi-Monte Carlo methods, I did not attend many talks but had a long run in the countryside (even saw a pheasant and a heron), worked at “home” on pressing recruiting evaluations and had a long working session with Pierre Jacob. Plus an evening out sampling (just) a few Belgian beers in the shade of the city hall…

Today was more in my ballpark as there were MCMC talks the whole day! The plenary talk was not about MCMC as Erich Novak presented a survey on the many available results bounding the complexity of approximating an integral based on a fixed number of evaluations of the integrand, some involving the dimension (and its curse), some not, some as fast as √n and some not as fast, all this depending on the regularity and the size of the classes of integrands considered. In some cases, the solution was importance sampling, in other cases, quasi-Monte Carlo, and yet other cases were still unsolved. Then Yves Atchadé gave a new perspective on computing the asymptotic variance in the central limit theorem on Markov chains when truncating the autocovariance, Matti Vihola talked about theoretical orderings of Markov chains that transmuted into the very practical consequence that using more simulations in a pseudo-marginal likelihood approximation improved acceptance rate and asymptotic variances (and this applies to aBC-MCMC as well), Radu Craiu proposed a novel processing of adaptive MCMC by treating various approximations to the true target as food for a multiple-try Metropolis algorithm, and Luca Martino had a go at resuscitating the ARMS algorithm of Gilks and Wild (used for a while in BUGS), although the talk did not dissipate all of my misgivings about the multidimensional version! I had more difficulties following the “Warwick session” which was made of four talks by current or former students from Warwick, although I appreciated the complexity of the results in infinite dimensional settings and novel approximations to diffusion based Metropolis algorithms. No further session this afternoon as the “social” activity was to visit the nearby Stella Artois brewery! This activity made us very social, for certain, even though there was hardly a soul around in this massively automated factory. (Maybe an ‘Og post to come one of those days…)

## sticky Metropolis

Posted in Statistics, University life with tags , , , , , , on September 6, 2013 by xi'an

My former student Roberto Casarin and his colleagues wrote (and arXived) a paper entitled Adaptive sticky generalized Metropolis algorithm. The basic idea is to use some of the rejected and past values of the chain to build an adaptive proposal, the criterion for choosing those values being related with the distance at the rejected point between the target and the proposal. In a sense, it gives a reward to surprising points, i.e. points where the proposal does poorly in approximating the target. On top of this, they include a multiple-try strategy where several values are generated from the current proposal and one of them is selected, to be accepted or rejected in a Metropolis step. The learning set may include several of the proposed (and rejected) values. This paper generalises Holden, Hauge and Holden (AoAP, 2009) and extends their proof of stationarity. The authors explore at length (the paper is 63 pages long!) the construction of the adaptive proposal distribution. This construction appears to be quite similar to Gilks’ and Wild’s (1993) ARMS algorithm. Hence, unless I missed a generalisation, it seems to me that the solutions are restricted to unidimensional settings. For instance, the authors propose to implement their algorithm for each complex conditional in a Gibbs sampler, meaning starting from scratch and running a large enough number of iterations to “reach” convergence. I also wonder at the correspondence between this construction and the original assumption of a minorisation condition wrt the target density in the event of an unbounded support. While this paper represents an interesting extension of the automated simulation algorithms of the ARMS type, and while the method is investigated thoroughly by several simulation experiments (in the second half of the paper), I remain somehow circumspect at the possibly of using ASMTM in complex high-dimensional problems as the learning cost soar with the dimension.

## adaptive Metropolis-Hastings sampling using reversible dependent mixture proposals

Posted in Statistics with tags , , , , , on May 23, 2013 by xi'an

In the plane to Birmingham, I was reading this recent arXived paper by Minh-Ngoc Tran, Michael K. Pitt, and Robert Kohn. The adaptive structure of their ACMH algorithm is based upon two parallel Markov chains, the former (called the trial chain) feeding the proposal densities of the later (called the main chain), bypassing the more traditional diminishing adaptation conditions. (Even though convergence actually follows from a minorisation condition.) These proposals are mixtures of t distributions fitted by variational Bayes approximations. Furthermore, the proposals are (a) reversible and (b) mixing local [dependent] and global [independent] components. One nice aspect of the reversibility is that the proposals do not have to be evaluated at each step.

The convergence results in the paper indeed assume a uniform minorisation condition on all proposal densities: although this sounded restrictive at first (but allows for straightforward proofs), I realised this could be implemented by adding a specific component to the mixture as in Corollary 3. (I checked the proof to realise that the minorisation on the proposal extends to the minorisation on the Metropolis-Hastings transition kernel.) A reversible kernel is defined as satisfying the detailed balance condition, which means that a single Gibbs step is reversible even though the Gibbs sampler as a whole is not. If a reversible Markov kernel with stationary distribution ζ is used, the acceptance probability in the Metropolis-Hastings transition is

α(x,z) = min{1,π(z)ζ(x)/π(x)ζ(z)}

(a result I thought was already known). The sweet deal is that the transition kernel involves Dirac masses, but the acceptance probability bypasses the difficulty. The way mixtures of t distributions can be reversible follows from Pitt & Walker (2006) construction, with  ζ  a specific mixture of t distributions. This target is estimated by variational Bayes. The paper further bypasses my classical objection to the use of normal, t or mixtures thereof, distributions:  this modelling assumes a sort of common Euclidean space for all components, which is (a) highly restrictive and (b) very inefficient in terms of acceptance rate. Instead, Tran & al. resort to Metropolis-within-Gibbs by constructing a partition of the components into subgroups.