Learning-NUM: Utility Maximization in Stochastic Queueing Networks