Online Deep Learning in Wireless Communication Systems

We consider the problem of optimal power allocation in wireless fading channels. Due to interference, this problem is non-convex and challenging to solve exactly. The resemblance of this problem to a statistical loss problem motivates the use of a learning parameterization of the power allocation function. In particular, we use deep neural networks (DNNs) to represent the power allocation and develop a primal-dual learning method to train the weights of the DNN. Because the channel and capacity models may not be known in practice, we extend the learning algorithm to permit stochastic online operation, in which gradients are approximated by sampling. We demonstrate in a series of numerical simulations the performance of the proposed online primal-dual learning method in training a DNN-parameterization relative to a well-known heuristic benchmark.