Control of prior distribution about octis HOT 3 CLOSED

Hi,
yes, skopt.space allows you to specify a prior over the numeric hyperparameters. The possible values are "uniform" and "log-uniform".
Thanks for your questions, we are going to make this clearer in the OCTIS documentation.
Cheers,
Bruno

from octis.

Hi, thank you for your interest.
For the Bayesian Optimization part, we rely on the scikit-optimize library. The ways to control the prior over BO are the following:

1. use different surrogate models, i.e. Gaussian Process and Random Forest, and this can be specified both in the python library and from the dashboard

2. choose a specific kernel for the Gaussian Process (this feature is available only in the python library at the moment). The default kernel is 1.0 * Matern(length_scale=1.0, length_scale_bounds=(1e-1, 10.0), nu=1.5), but you can pass any other kernel from sklearn (See "sklearn.gaussian_process.kernels" here: https://scikit-learn.org/stable/modules/classes.html).

Hope this helps!

from octis.

Thanks for the prompt reply!

I now see that control on surrogate model is possible. What I also meant is for example in hyperparameter optimization, one usually want to be able to control the distribution (normal, log-normal, uniform, log-uniform etc.) to sample from for each hyperparam separately. skopt.space classes (e.g. Real, Integer) has some control on this if I remember correctly.

from octis.