Closed rbarbudo closed 3 years ago
It sounds like you're on the right track. Unfortunately Auto-WEKA doesn't directly process the parameter space, but hands it to SMAC to do that. So what you have in mind wouldn't be a trivial task because you would have to assemble the parameter space from the configuration files and take dependencies into account yourself.
I wonder whether it would be possible to use a random search for exploring the Auto-WEKA search space. More specifically, I would like to randomly generate pipelines consisting in two elements (feature selection and classifier/regressor) with the same WEKA algorithms and hyper-parameter space used with the SMAC algorithm.
As far I understand, it would be necessary to create an ExperimentConstructor (I think an autoweka.ListExperimentConstructor) similar to the autoweka.GridSearchExperimentConstructor. Then, this ExperimentConstructor should be instantiated within the weka.classifiers.meta.AutoWEKAClassifier. Am I wrong?
I would really appreciate if you could give me any insight on how to develop this randomized version of Auto-WEKA. Thank you in advance for your time and the Auto-WEKA project.