About speed, epsilon and wait

Hi all,
I'm trying to figure out a decent set of parameters to achieve a good trade-off between speed and accuracy.
On my data, default parameters make comuputation times extremely long (especially compared to other community finding approaches), but we have noticed looking at the log that even 100 consecutive iterations without record breaking are quite rare, would it be reasonable to set wait=100, possibly increasing nbreaks, then?
Also, are "wait" parameter and "niter" passed to mcmc_args somehow related, in the sense that it could be opportune to increase one and decrease the other or should be they considered independent? Reading the older posts I've found there's an "epsilon" parameter which could be tuned in order to converge faster, and we are definitely trying to increase it (1e-3, to start). I've noticed another epsilon in hierarchy_minimize() parameters, is that expected to influence speed as well?
I may have additional questions on specific parameters (such as d and c) but I guess I can save them for another post.
Thanks

d

Hi all,
I'm trying to figure out a decent set of parameters to achieve a good trade-off between speed and accuracy.
On my data, default parameters make comuputation times extremely long (especially compared to other community finding approaches), but we have noticed looking at the log that even 100 consecutive iterations without record breaking are quite rare, would it be reasonable to set wait=100, possibly increasing nbreaks, then?

This is difficult to say in general, as it depends on the properties of
the posterior distribution. But often what you say is correct: it's
sufficient to keep one of either nbreaks or wait fixed, and only vary
the other.

Also, are "wait" parameter and "niter" passed to mcmc_args somehow related, in the sense that it could be opportune to increase one and decrease the other or should be they considered independent? Reading the older posts I've found there's an "epsilon" parameter which could be tuned in order to converge faster, and we are definitely trying to increase it (1e-3, to start). I've noticed another epsilon in hierarchy_minimize() parameters, is that expected to influence speed as well?

niter just controls how many iterations are performed for each sweep,
and wait controls how many sweeps are performed.

The difference between the epsilon in mcmc_equilibrate() and in
hiearchy_minimize() is explained in the documentation.

Hello Tiago,
thanks for the answer

No, it will not. From the documentation:

    epsilon: ``float`` (optional, default: ``1e-8``)
        Only replace levels if the description length difference
        is above this threshold.

This is about numerical precision, nothing else. Most of the time you
should leave this parameter alone.