Annealing scheme for lambda

When lambda is sampled from the posterior distribution right from the start, mixing and convergence of the Markov chain might be prohibitively slow. The program therefore offers an annealing scheme. During the burn-in period, lambda is sampled from a mixed distribution, where the mixing parameter increases linearly from 0 (pure prior) to 1 (pure posterior). There are two options for mixing the distributions. One can either mix the parameters of the distribution, or one can mix the distributions themselves. The following figures demonstrate the difference between these mixing schemes. The numbers over the subfigures show the respective value of the mixing parameter.

Annealing by mixing parameters

You select this annealing scheme with the following choice in the run settings menu:

Annealing scheme for lambda: PAR

Annealing by mixing parameters

Annealing by mixing distributions

You select this annealing scheme with the following choice in the run settings menu:

Annealing scheme for lambda: PROB

Annealing by mixing probability distributions

Obviously, mixing the distributions rather than the parameters results in a distribution with longer tails, which can be advantageous to leave a metastable state during equilibration.


Last modified: July 2001

Back to the beginning of this documentation.