This is at a high level how one of our Google internal black box optimizers behaves. The term used is "infeasible region" but it's the same idea as using a nan.
I would caution against using nan to always mean infeasible. Instead users should catch experiments outside the feasibility region and return a special infeasible value. This will increase visibility into the behavior of the optimizer, because it leaves nan to be used for values inside the region of constraint that are still problematic (due to bugs, numerical instability, etc)
I would caution against using nan to always mean infeasible. Instead users should catch experiments outside the feasibility region and return a special infeasible value. This will increase visibility into the behavior of the optimizer, because it leaves nan to be used for values inside the region of constraint that are still problematic (due to bugs, numerical instability, etc)