Re: Acceptance rate limits for mcmc in dice

[Eva Ntormousi]

Hello everyone,

I just started using dice, and I have a small question regarding the default acceptance rate limits for the mcmc algorithm.
It was my understanding that the asymptotic acceptance rate for a mcmc should be 23.4% for a multi-dimensional parameter space.  However, the default values in dice are between 80-90%.  Is this because of a difference in notation, or due to a variation in the method?

Thank you very much in advance for your answers!
Cheers,
Eva

2017-06-19 18:21 GMT+03:00 Eva Ntormousi <ent...@gmail.com>:

Hello everyone,

I just started using dice, and I have a small question regarding the default acceptance rate limits for the mcmc algorithm.
As far as I know the asymptotic acceptance rate for a mcmc should be 23.4% for a multi-dimensional parameter space.  However, the default values in dice are between 80-90%.  Is this because of a difference in notation, or due to a variation in the method?

Thank you very much in advance for your answers!
Cheers,
Eva

[Valentin Perret]

Hi Eva,

I wasn't fully aware of the 23.4% magic value. However I observed through several tests that a high acceptance (but lower than 100%) is always preferable to lower the noise. Maybe this makes the convergence slower, but for the quality of the initial conditions my feeling is that this is better.

Note also that the MCMC is implemented with a multi-try version of the algorithm (if you set mcmc_ntry>1), which is supposed to optimise the acceptance rate regardless of the step function. You can try with the 23.4% value and let me know how it goes!

Cheers,

Valentin

[Eva Ntormousi]

Hi Valentin,

 For some reason I was not notified of your answer... 

Thank you for the tip! In the meantime I have indeed run tests with different acceptance fractions, and indeed it made little difference, as long as it was not too high.

Cheers,

Eva