It seems like fancy hyperparameter optimization techniques (e.g. Bayesian black-box optimization) probably don't help here either, because they don't solve the problem of extrapolating <i>outside</i> the range of hyperparameter values have have already been tried. Is that a valid conclusion?