Extreme Learning Machine (ELM) for Fast User Clustering in Downlink Non-Orthogonal Multiple Access (NOMA) 5G Networks

Non-orthogonal multiple access (NOMA) has been envisaged as a successor of orthogonal multiple access (OMA) in the fifth generation (5G) networks and beyond because NOMA has been theoretically and empirically proven to be more bandwidth-efficient than OMA. Nevertheless, user clustering (UC) in NOMA is another prevalent issue. To maximize the throughput and fulfill the successive interference cancellation (SIC) constraints, the UC has been formulated as a clustering optimization problem which has been extensively researched in literature. Recently, an artificial neural network-based UC (ANN-UC) scheme has emerged as a viable solution that can optimally cluster users after exhaustive training. However, the ANN model has an extremely slow learning speed, due to the gradient-based back-propagation (BP) algorithm used by the ANN. To address these issues, this paper proposes a novel fast-learning extreme learning machine-based UC (ELM-UC) scheme. Unlike the ANN-UC technique, the input weights and the bias for the hidden layer nodes of ELM are randomly generated and tuning of parameters is not required, thereby leading to a faster learning rate. In this work, the ELM architecture is adapted to operate in NOMA environments where the optimal cluster formation can be predicted rapidly based on the users’ channel gains and powers. Performance comparisons with the state-of-the-art UC schemes are investigated via extensive simulations. Remarkably, simulation results demonstrate that the proposed ELM-UC technique can achieve near-optimal performance compared to the brute-force search (B-FS) method and outperforms the existing clustering techniques including ANN-UC and dynamic user clustering (DUC).