search

Yelp / MOE

A global, black box optimization engine for real world metric optimization.
Other
1.31k stars 140 forks source link

[Question] Support for integer/categorical data #369

Closed yonromai closed 10 years ago

yonromai commented 10 years ago

Hi,

Gp are supposed to work on smooth parameter spaces. Are there any plans to add support for integer and categorical parameters to MOE?

Thanks! Romain

sc932 commented 10 years ago

Hi Roman,

There are plans to extend the Bayesian Global Optimization framework to include non-continuous spaces. Right now you can kind of "fake it" in two ways, or use our bandits framework:

  1. Pretend the space is continuous and then project the output MOE gives onto your constrained domain (like if MOE suggests to sample the point [2.04, 14.8, 0.003] and the first coordinate needs to be restricted to the naturals you can take the floor and sample [2, 14.8, 0.003] instead)
  2. If you have a small number of categories, or they are independent, you can run MOE independently for each and then just sample the points of highest expected improvement across all categories.
  3. If you have a small number of categorical parameters and no continuous parameters you should check out our Multi-Armed Bandit solutions (coming in v0.2.0, should be done today/tomorrow).

I'll close this issue for now. Please shoot me an email at sclark@yelp.com with any other questions and I'd be happy to help however I can!

Cheers, Scott