A Neural Network based Power Allocation Algorithm for D2D Communication in Cellular Networks

This paper studies the power allocation problem in device-to-device (D2D) communications underlaying cellular networks. A Q-learning based distributed power allocation (Q-PA) algorithm is first proposed, which attempts to obtain optimal power allocation through iterative updating of Q-value tables. Based on the Q-PA algorithm, a neural network (NN) based power allocation (N-PA) algorithm is further proposed, in which training data obtained using the Q-PA algorithm are used to train an NN model, and the trained NN model is used to perform power allocation for D2D users. Simulation results show that both the Q-PA algorithm and the N-PA algorithm can improve the system throughput as compared with two traditional algorithms. The N-PA algorithm can significantly reduce the time cost for power allocation at a small expense of the system throughput as compared to the Q-PA algorithm.