Kyle Beauchamp

Here are some things that I didn't implement: Estimation with a fixed value of equilibrium populations. Sparse estimation. Estimation with a fixed sparsity structure. These are considerably more difficult, but...

So I have been experimenting with NumExpr for speeding up likelihood calculations (this is for my ensemble stuff, but it would also apply here and to MLE code in MSMB)....

Similar, but not identical. I think pymc uses normally distributed proposal steps, with a proposal scale that is tuned to optimize the acceptance rate. In Noe's paper, he analytically calculates...

Well, there's also the question of prior. For a general transition matrix, the dirichlet prior is pretty standard. I don't think it's as clear-cut how to parameterize the prior for...

I think there's a lot that could be done to generalize and accelerate this code, but I think it might be wise to wait to see _how_ we plan to...

Regarding speed, there are several things that might be rate limilting, depending on what regime we work in: 1. Large numbers of pymc `Potential` function calls (fix with Cython?) 2....

OK so it appears that I will likely need to rewrite this code to better handle NANs from nonzero counts.

_zero_ counts, rather.

Argh. So I think there's a bug in pymc that won't let you initialize a Dirichlet RV at a boundary point. If you have alpha=1 (e.g. uniform), then that point...

So do you want this as a general trajectory analysis tool or for MSM clustering. Regardless, this feature should be in MDTraj. On Wed, Apr 2, 2014 at 2:49 PM,...