Pre-optimization and adaptation

[florianhartig]

It would be nice to have a more general concept for pre-optimization and adaptation.

At the moment, only the Metropolis MCMCs do pre-optimization. The first thing I'd like to look at is we can't generalize this to all MCMCs (setting start values etc). This is also connected to

https://github.com/florianhartig/BayesianTools/issues/28 https://github.com/florianhartig/BayesianTools/issues/79

[TankredO]

My idea would be this:

• remove the coupling between setupStartPropsal and the pre-optimization for Metropolis
• create a function preOptimize (or something like that) which takes a bayesianSetup and a target sampler and returns a startValue data type
• call preOptimize in runMCMC for the corresponding sampler when settings$optimize == TRUE
• overwrite settings$startValue with startValue object
• start the sampler with startValue

The problem is that when using a new data type / object for start values we need to change the code that handles start values in all samplers.

If we want to keep changes to the code to a minimum we could also return different objects for the different preOptimize targets. E.g. start values and covariance matrix for Metropolis, a 3 * nPar matrix of start values for DE, ... .

What do you think? @MaximilianPi @florianhartig

[florianhartig]

stale branch https://github.com/florianhartig/BayesianTools/tree/issue87 connected with this - check if any of this should be merged

[florianhartig]

Not sure this can be done for 0.1.7, because potentially complicated, but @JohannesOberpriller, have a look at what's done so far