search

HIPS / Spearmint

Spearmint Bayesian optimization codebase
Other
1.55k stars 327 forks source link

Current implemented features #46

Open jmgo opened 8 years ago

jmgo commented 8 years ago

Hi!

I've read the papers about spearmint and i got to say that there seems to be a lot of good stuff in there.

However, I noticed that part of that stuff is not implemented in the current version. I think that a simple list of things that are not currently implemented would help people who see the code for the first time.

From what I've seen these are the things that are not implemented (mainly from the paper 'Multi-task bayesian optimization'):

Can you confirm me that these features are not implemented?

Also when I set max-concurrent jobs to >1, is somekind o parallelization strategy performed? Perhaps the one presented in the paper 'Practical Bayesian Optimization of Machine Learning Algorithms'? "resources" : { "my-machine" : { "scheduler" : "local", "max-concurrent" : 2, "max-finished-jobs" : 100

Best Regards, Jorge

JasperSnoek commented 8 years ago

Hi @jmgo, yes that's right. The various innovations in the Multi-Task Bayesian Optimization paper are not currently implemented here. Other acquisition functions are also not available as far as I know in the master branch although some newer Predictive Entropy Search algorithms are implemented in the PESC (https://github.com/HIPS/Spearmint/tree/PESC) branch.

The parallelization implementation is the one described in the "Practical Bayesian Optimization" paper. We integrate over the possible outcomes of the currently running experiments.

I hope that helps!

Jasper

jmgo commented 8 years ago

Hi, Jasper!

Thanks for the info, it was exactly what I wanted to know. Also, thanks for the tip for the PESC branch, I didn't check the branches.

P.S: Please mention in the readme that the features Multi-Task Bayesian Optimization are not implemented. I was looking for them for while in code.

Best Regards, Jorge