Active Learning for Function Approximation