Problems involving physical/tangible systems (such as production scheduling, crew assignment, ...) are inherently constrained, whether the modeler bothers to put all the constraints in or not. Any variable representing a physical process has a finite domain (presuming that the physicists are correct in saying there is a finite amount of matter in the known universe), whether the model author bounds them or not.
On the other hand, consider the problem of fitting a regression model (least squares) using gradient descent. The variables are the model coefficients, and while there may be ways to argue that this particular coefficient cannot be astronomically large in magnitude, it's unlikely there will be any known a priori bounds for the coefficients. (The computer's limited range of floating point values will inherently bound them, whether you like it or not.) So the problem is treated as unconstrained, with no loss of validity or functionality.