Can I restart an MCMC process from a previous one?

Sometimes I run a model with brm() for a few days or weeks, only to realize that there is some convergence problem as seen from one of the Rhat’s being greater than 1.1 or too small an effective sample size. If I have already saved the .RData file from a previous MCMC process, can I restart it somehow by setting longer chains to achieve better convergence or a higher effective sample size? One issue is that I cannot change the the number of warm-up (or “burn-in”) iterations?

You can’t. There have been multiple threats about such a feature already and if this will ever come it has to be implemented in rstan or even Stan itself.

While you cannot restart things… what takes most of the time during warmup is finding the mass matrix (or better, the diagonal of it). The diagonal of the mass matrix is serialized and with cmdstan you have the option to use it again. However, I have to say that this works technically, but it is not very convenient to do… but it does work and I found this to be very useful in some examples.

1 Like