With (imho closely related) bayesian statistics you can add those rules into your model, the problem is that it quickly becomes computationally intractable to d

With (imho closely related) bayesian statistics you can add those rules into your model, the problem is that it quickly becomes computationally intractable to do exactly (in a calculate this N dimensional integral with N>>1000 kind of way). So in the end you still end up fitting some simplified vaguely similar model/function to your data and/or using sampling to keep things tractable.