With SVM, you often must perform a rather larger grid search over kernels and kernel parameters. It seems like no matter the model, we can't avoid the hyperpara

With SVM, you often must perform a rather larger grid search over kernels and kernel parameters. It seems like no matter the model, we can't avoid the hyperparameter problem -- although boosting and bagging meta-methods come close. It would be nice if we could quantify the complexity of a dataset and match this to a model with similar complexity. I imagine that it's hard (or impossible) to decouple these two complexity quantifiers, however.