Sequential Monte Carlo on large binary spaces

Christian Schäfer and Nicolas Chopin (both from CREST) just posted a new version of their paper, Sequential Monte Carlo on large binary sampling spaces, where they construct adaptive sequential algorithms for running variable selection with a large number of variables. The optimisation of their Gibbs scheme is thus over distributions exploring the set of the 2^d possible configurations of the variable space, using a logistic regression to decide on the inclusion of the variables. The performances of the method are quite convincing! (The above is comparing their method (gold) with AMCMC (blue) and regular MCMC (red).) The code is available in Python.