User scheduling for uplink OFDMA systems by deep learning

User scheduling is an efficient way to harvest the frequency and multiuser diversity gain for uplink Orthogonal Frequency Division Multiple Access (OFDMA) system. To solve the non-convex scheduling problem, existing numerical or searching based solutions face the difficulty of meeting the real-time requirement of fast scheduling. In this paper, a deep learning based method is proposed to solve the user scheduling problem, aimed at reducing the scheduling complexity for real-time implementation. The key challenge of learning the scheduling decisions lies in how to ensure that the learned decisions satisfy the coupled binary constraint. To tackle the difficulty, we design a deep neural network (DNN) to approximate the binary vector quantization operation. The DNN is then used as the activation function in the output layer of another DNN, where the latter is trained to directly maximize the performance utility via unsupervised learning. Simulation results demonstrate that the proposed method is able to largely reduce the complexity with marginal performance and fairness loss compared to the greedy searching method.