Running summarizing tools on a new article

Abstract: We consider the optimal power allocation with a set of transmitter/receiver pairs in a large scale wireless network. Obtaining an optimal policy requires the solving of a non-convex, infinite dimensional optimization problem, motivating the use of a learning model parameterization of the optimal policy. Both for its theoretical properties and experimental performance, deep neural networks are a strong choice to parameterize the power allocation policy. However, fully connected neural networks are impractical to both train and implement for large scale networks, as their size grows quickly with network size. We propose using a convolutional neural network architecture, specifically one that employs a graph convolution over the randomly varying fading links of the wireless network. Such random edge graph neural networks (REGNNs) are shown to retain an important permutation invariance property of the optimal power allocation. Numerical simulations show strong performance of a primal-dual learning method in training an REGNN parameterization relative to existing heuristic benchmarks.