Learning the Value of Approximation