0Deep Learning Based Power Control for Cell-Free Massive MIMO with MRT

Cell-Free Massive MIMO with MRT (MaximumRatio Transmission) has the advantage of decentralized beamforming with the smallest front-haul overhead. Its downlink power control plays a dual role of fair power distribution among users and interference mitigation. It is well-known that finding the optimal max-min power control relies on SOCP (Second Order Cone Programming) feasibility bisection search, whose large computational delay is not suitable for practical implementation. In this paper, we devise a deep learning approach for finding a practical near-optimal power control. Specifically, we propose a convolutional neural network that takes as input the channel matrix of large-scale fading coefficients and outputs the total transmit power of each AP (access point). Using this information, the downlink power control for each user is then computed by a low-complexity convex program. Our approach requires to generate far fewer training examples than existing schemes. The reason is that we augment the training dataset with magnitudes larger number of artificial examples by exploiting the special structure of the problem. The resulting deep learning model not only provides a near-optimal solution to the original problem, but also generalizes well for problems with different number of users and different propagation morphologies, without the need to retrain it. Numerical simulations validate the near optimality of our solution with a significant reduction in computational burden.