At the moment MCHMC/MCLMC makes sure it targets the correct distribution by controlling the error when solving the SDE's of the Langevin mechanics. This is very fast. However, tuning the target error can be quite hard since it is a quantity between [0, inf]. We would like to add the option to metropolis adjust the sampler as a way of tuning the acceptance probability. This is easier than tuning the energy error since the acceptance probability is a number between [0, 1], usually 0.6 or 0.9.
Background
At the moment MCHMC/MCLMC makes sure it targets the correct distribution by controlling the error when solving the SDE's of the Langevin mechanics. This is very fast. However, tuning the target error can be quite hard since it is a quantity between [0, inf]. We would like to add the option to metropolis adjust the sampler as a way of tuning the acceptance probability. This is easier than tuning the energy error since the acceptance probability is a number between [0, 1], usually 0.6 or 0.9.
How to do this:
Stepmethod implement this codeTAPwhich controls the acceptance.TAPis provided add a tuning routine in the tuning module and cache the associated TEV.