Passing dotdotdot to likelihood function about bayesiantools

[florianhartig]

hmm .... I guess why not, that should be possible.

We would have to discuss how we handle restarts, i.e. can the user change this setting if the MCMC is restarted or should we save it, and once it was set for the first time, it cannot be changed any more.

I'm putting this up as a possible enhancement, not sure when we'll find the time to implement this.

In the meantime, I guess you can solve your problem with a small wrapper as in

runMyMCMC <- function(alpha, beta, ... ){
likelihood <- function{
...
alpha, beta
}
runMCMC (likelihood, ...)

}

[florianhartig]

Hi Dominik,

glad you find BT helpful!

About the ... -> what would you like to pass to the likelihood?

We discussed whether it would be sensible to let the user pass the data, but then the MCMC needs to store the data somewhere, plus this would break compatibility with optim and sensitivity, so we decided against it.

[dleutnant]

Thanks for your quick response!
In my case, the likelihood function calls an external model which basically follows the following structure (in magrittr jargon)

LL <- function(x) {
 run_external_model(param = x) %>%
  evaluate_model(.) %>%
  do_post_processing(., alpha = 1.234, beta = 5.678) %>%
  calculate_loglikelihood(.)
}

Here, the dummy parameters alpha and beta are individually set according to the calibration setup and are fixed during the MCMC sampling. So in my case, the following would be ideal:

LL <- function(x, ...) {
 run_external_model(param = x) %>%
  evaluate_model(.) %>%
  do_post_processing(., list(...)) %>%
  calculate_loglikelihood(.)
}

With this approch, I could easily create multiple MCMC setups for further analysis. stats::optim offers a comparable ... "which passes further arguments to fn and gr".

[florianhartig]

I see. And then you would want to provide the ... to the runMCMC function I guess?

[dleutnant]

exactly!

[dleutnant]

Cool! Thanks! Let me know if I can support you with this (e.g, testing, etc)!

[florianhartig]

OK, guys, I think we should make a decision about this ... I don't want to drag this into the next release.

I'm kind of leaning against implementing this, as it doesn't seem to be a huge problem for the user to simply change the likelihood, or, alternatively (a bit dangerous though), let the likelihood access a global variable that is then changed.

Providing additional comments in the rumMCMC commands would mean passing them from the rumMCMC to the samplers, to the likelihood calls ... it's not a huge effort, but it's more code that we have to maintain.

Additionally, if we do stuff like the WAIC calculations or other things, the parameters need to be available, so they would need to be stored somewhere.

Once the stuff is stored, the issue arises that a user might re-use a setup with stored additional parameters in a different context, where they modify again the parameters, in which case we have to throw a warning or whatever.

It all seems to me as if we create a pretty fragile structure for a small improvement in convenience.

[dleutnant]

That's fine for me, although I think being in line with other parameter estimators or samplers (stats::optim, DEoptim::DEoptim, mco::nsga2, adaptMCMC::MCMC) promotes users to feel comfortable at first sight.
However, I can't assess additional efforts required to support this as I am not that deep into the code basis.

[florianhartig]

OK, this is closed