Closed ihincks closed 7 years ago
Thinking about it again, I think I have est_update_covariance
wrong in the not-fixed-covariance case. I think I should be taking the mean over particles after computing the covariance matrix for each particle. Do you agree?
For the docstring as written, I agree, est_update_covariance
should take the mean after computing the covariance matrix. That said, it's also an operationally-relevant question to ask what the covariance for the estimated noise parameters are; if I understand correctly, that's the question answered by the current implementation of est_update_covariance
?
I think that should do it for est_update_covariance
. I'm not exactly sure what the previous version corresponded to. It could correspond to what you are talking about, but it's not clear to me. In any case, all I wanted this function to do is to fill the gap of SMCUpdater.est_mean
for the new model parameters, when one actually cares about the covariance matrix instead of the things that parameterize it.
Any further concerns/requests on this PR?
My apologies for falling behind. Anyway, I think this looks good so I'll go on and merge it in now. Thanks!
This class is similar to
RandomWalk
except that it only supports zero-mean multivariate normal distributions, and it has the ability to automatically add model parameters which parameterize the covariance matrix, so that they can be learned in tandem with the parameters of interest.See
doc/guide/timedep.rst
for a demo, which is worth more than what I can write here.