search

XanderHorn / autoML

Automated machine learning in R
MIT License
9 stars 7 forks source link

exploiting bayesian optimization from mlrHyperopt? #1

Open KnutJaegersberg opened 5 years ago

KnutJaegersberg commented 5 years ago

I like the integrated approach of your autoML package. Can optimization be improved (lower error with less training budget) compared to random tuning and frace optimization by including mlr hyperopt / mlrMBO as an option? In mlrhyperopt, there is access included to a couple of standard tuning ranges according to their user fed hyperparameter "database", for several standard models. http://mlrhyperopt.jakob-r.de/parconfigs

I have used mlrhyperopt from time to time and found the results I got from that hyperparameter tuning api quite useful.

XanderHorn commented 5 years ago

Whilst having used mlrHyperOpt to generate most of the hyper parameter grids internally in the package, I do not use it for all models.

Bayesian optimisation can be included, I excluded it due to installation issues when testing the package, quite quick to implement again. At the moment Iterated f-racing is also included if random tuning does not suffice.