A Deep Neural Architecture for Real-Time Access Point Scheduling in Uplink Cell-Free Massive MIMO

Abstract

In this paper, a novel hybrid architecture is proposed combining expert knowledge for optimal power allocation and deep artificial neural networks (ANN) to address the access-point scheduling problem in cell-free massive multiple-input multiple-output (MIMO) communication systems with a serial bandwidth-limited fronthaul architecture. The scheduling task is formulated as an image segmentation problem for which a supervised encoder-decoder like ANN is proposed. It consists of serially concatenated contraction and expansion layers to maximize the (regularized) cross-entropy, followed by a binary projection to undo the relaxation problem. Besides the robustness to scenarios with a time-varying system load and fronthaul bandwidth, the proposed architecture provides a complexity-efficient solution that fulfills the fronthaul bandwidth constraints and satisfies the real-time considerations. Our experimental results verify the competitive performance of the proposed solution with respect to both nonlinear solvers and state-of-art convex algorithms while the time efficiency of the ANN model outperforms the state of the art, especially, in scenarios with a large number of users.