Optimization of categorical parameters

[ogencoglu]

Thanks for the great work of pybads. Quick question: Can we optimize both continuous and categorical parameters somehow? Does it provide any capability for categorical variables?

[lacerbi]

Thanks for your interest, and great question!

tl;dr answer: at the moment PyBADS (and similarly the underlying BADS method) does not support categorical variables.

In the near- to medium- future we could perhaps add support for integer variables via some tricks shown in the Bayesian optimization literature.
You could then map any discrete (ordered) set to a range of integers.
However, this trick would not work particularly well if your variables are truly categorical (i.e., there is no clear ordering, and the function takes completely different shapes based on the categorical value).

Implementing truly categorical variables would be a very interesting future project.

[ogencoglu]

Thanks for the swift reply!