Bayesian Model Selection (BMS) is a technique similar to what you’ve described
In BMS, the model itself is considered as a parameter in the model.
i.e. at each iteration, instead of just sampling possible parameter values, it changes the actual parameters included in the model.
Techniques for accomplishing this include Reversible Jump MC.
Here’s an article about it in R:
It’s possible to set up in some form in pymc3, in fact I think I’ve seen some exmaples, but there are various problems/difficulties to work through to use it