The fastest implementation of garli

[rob.l...@gmail.com]

Hi Derrick,

A quick question. We are planning a big project to analyse about 50 of datasets in garli, and I'm wondering what the fastest version is. The datasets will be partitioned, and we'd like to do ~1000 bootstraps on each. They are mostly DNA datasets.

I took a look at the google code site, and noticed that you have a separate 'beagle' branch. Given the 2012 paper on BEAGLE and garli, I thought I'd ask if this is the best way to go?

Cheers,

Rob

[Derrick Zwickl]

Hi Rob,

No, the Garli Beagle version won't do what you want.  I implemented that prototype version on top of Garli 1.0, but numerical precision issues turned out to be a big headache.  So, it hasn't been updated and there is no partitioned Garli version that uses Beagle.  The numerical precision turns out to be a big issue for ML inference because the likelihood surface has to be flat to optimize on it, while in a Bayesian context a little scoring inaccuracy that causes a wavy surface ends up not being much of an issue.  So, Beagle is really great for the Bayesian programs.

So, there is probably not a particular Garli version that would be fastest for you, although I could probably give some recommendations on settings that would be the fastest and most appropriate for your particular datasets.

Cheers,
Derrick