On Sun, May 2, 2010 at 3:59 PM, Chou-Ching K. Lin <[log in to unmask]> wrote:
> Dear Vladimir:
> Thank you for your help.
> I think I will follow your suggestion for running DCM_IR in batch.
>> 2. Last time, I asked about the magic number of 64 in DCM_IR. I did change
>> it to larger numbers and it can run longer. My question is that when it
>> stops and does not converge. It you run again with existing priors, I found
>> the F starts from near 0, instead of continuing the existing number.
>>Did you answer 'yes' to both questions 'Use previous priors?' and 'Use
>>previous posteriors'? If not, you should.
> Yes, I answered yes to both questions.
>> My question is that, for the final F, is the F of previous 64 runs stored in
>> the saved file or F is re-calculated?
These are not 64 'runs' but 64 iterations of the optimization
procedure. With each iteration you should improve the model fit and
get a higher value of F. But there is not much use in keeping F values
from previous iterations. These are just intermediate results.
>>What's important is not the value of F but the parameter values at the
>>end of previous optimization. They should be used if you say 'yes' to
>>'use previous posteriors'.
> I thought F, being the variational free-energy, is used for BMS?
You are right but what I meant is that it's not enough (and not even
necessary) to know the value of F to be able to re-initialize the
inversion from the point where it stopped before. What you really need
to know is the posterior parameter distribution from the previous run.
>> And can I use previous priors if I
>> changes the weights, like A or B?
>>You can, but it's not necessarily meaningful in this case. I don't see
>>a reason to do it.
> I used previous priors and posteriors for similar models because sometimes it really takes very long to converge if run from new prior and posterior. And I don't know whether there is other ways to change the initial conditions. But sometimes they produced different answers.
What you can do is take one 'template' model for which you get a
decent fit and use it to initialize all the other inversions (using
the 'initialize' button). That will ensure that the solutions you get
are more or less in the same part of the parameter space and should
solve the problem of slow convergence. It'd be even better if you
choose as a template something that you won't later include in the
model comparison. When you optimize F value it shouldn't be 0 at the
beginning but can be any number (depending on the priors). What should
be true is that if you use different initializations the value of F a
the beginning should be different.