Today, I took part in the thesis defence of Amandine Shreck at Telecom-ParisTech. I had commented a while ago on the Langevin algorithm for discontinuous targets she developed with co-authors from that school towards variable selection. The thesis also contains material on the equi-energy sampler that is worth mentioning. The algorithm relates to the Wang-Landau algorithm last discussed here for the seminars of Pierre and Luke in Paris, last month. The algorithm aims at facilitating the moves around the target density by favouring moves from one energy level to the next. As explained to me by Pierre once again after his seminar, the division of the space according to the target values is a way to avoid creating artificial partitions over the sampling space. A sort of Lebesgue version of Monte Carlo integration. The energy bands
require the choice of a sequence of bounds on the density, values that are hardly available prior to the simulation of the target. The paper corresponding to this part of the thesis (and published in our special issue of TOMACS last year) thus considers the extension when the bounds are defined on the go, in a adaptive way. This could be achieved based on earlier simulations, using some quantiles of the observed values of the target but this is a costly solution which requires to keep an ordered sample of the density values. (Is it that costly?!) Thus the authors prefer to determine the energy levels in a cheaper adaptive manner. Namely, through a Robbins-Monro/stochastic approximation type update of the bounds,
My questions related with this part of the thesis were about the actual gain if any in computing time versus efficiency, the limitations in terms of curse of dimension and storage, the connections with the Wang-Landau algorithm and pseudo-marginal approximations, and the (degree of) likelihood of an universal and automatised adaptive equi-energy sampler.