Neural Network Aided User Clustering in mmWave-NOMA Systems With User Decoding Capability Constraints

Abstract

This paper proposes a computationally efficient two-stage machine learning based approach using neural networks to solve the cluster assignment problem in a millimeter wave-non orthogonal multiple access (mmWave-NOMA) system where each user's individual successive interference cancellation (SIC) decoding capabilities are taken into consideration. The artificial neural network (ANN) is applied in real time to assign users to clusters taking each user's instantaneous channel state information (CSI) and SIC decoding capabilities as inputs. The algorithm is trained offline on cloud resources, i.e., not using the base station (BS) compute resources. This training is done using a dataset obtained by offline computation of input parameters using the optimization algorithms called NOMA-minimum exact cover (NOMA-MEC) and NOMA-best beam (NOMA-BB) from our earlier work in this area. As a result, we term the proposed algorithms in this paper as ANN-NOMA-MEC and ANN-NOMA-BB, respectively. The problem with applying optimization techniques, even low-complexity heuristics, in live networks for user clustering is that they require a very large number of computation steps to make a clustering decision. If these clustering decisions are based on the instantaneous channel of hundreds of users, it becomes prohibitively complex to implement in practical systems on a millisecond granularity as required by beyond 5G (B5G) systems. Instead, our proposed approach takes all this complexity offline and even off the BS compute resources and instead only applies a trained neural network to make such clustering decisions at a microsecond granularity on hundreds of users. Simulation results show the effectiveness of the ANN-NOMA-MEC and ANN-NOMA-BB schemes as the neural network trained on offline simulation data performs comparably with the NOMA-MEC and NOMA-BB heuristics that is applying computationally intensive algorithms to make every clustering decision in a live network.