A Bipartite Graph Neural Network Approach for Scalable Beamforming Optimization

Deep learning (DL) techniques have been intensively studied for the optimization of multi-user multiple-input single-output (MU-MISO) downlink systems owing to capability handling nonconvex formulations. However, fixed computation structure existing deep neural networks (DNNs) lacks flexibility with respect system size, i.e., number antennas or users. This paper develops a bipartite graph network (BGNN) framework, scalable DL solution designed multi-antenna beamforming optimization. The MU-MISO is first characterized by where two disjoint vertex sets, each which consists transmit and users, are connected via pairwise edges. These interconnection states modeled channel fading coefficients. Thus, generic process interpreted as task over weighted graph. approach partitions procedure into multiple suboperations dedicated individual antenna vertices user vertices. Separated operations lead calculations that invariant size. realized group DNN modules collectively form BGNN architecture. Identical DNNs reused at all users so resultant becomes flexible Component trained jointly numerous configurations randomly varying sizes. As result, can be universally applied arbitrary systems. Numerical results validate advantages framework conventional methods.