In the same way that number of parameters or degrees of freedom can be used as a proxy for the potential of a regression model to learn complex relations - is there a measure that can be compared on the same scale between various model architectures including linear models, tree-based models, Gaussian process, boosting, ensembles etc?
Measures such as AIC and BIC require number of parameters, which are not universally comparable across different model types. The Vapnik–Chervonenkis (VC) dimension or generalised degrees of freedom would be appropriate (although open to other suggestions), although their utility appears more conceptual than computable.
Is there a method to quantify the potential of model to learn complex relationships in data that is comparable between all model types? And is this implementable in Python?