Active Learning for Approximation of Expensive Functions with Normal Distributed Output Uncertainty