Deep Learning-Based Energy Efficient Resource Allocation for Underlay Cognitive MISO Interference Channels

Abstract

In this paper, we investigate a deep learning (DL)-based resource allocation strategy for an underlay cognitive radio network with multiple-input-single-output interference channels. The beamforming vector and transmit power of secondary users (SUs) are optimized to maximize the sum energy efficiency (EE) of the SUs whilst maintaining quality-of-service for all transmissions, i.e., regulating the interference caused at the primary user to less than a given threshold whilst guaranteeing a minimum requirement for the spectral efficiency of each SU. To this end, a novel DL framework is proposed in which the resource allocation strategy is approximated by a well designed deep neural network (DNN) model consisting of three DNN units. Moreover, an efficient training methodology is devised, where the DNN model is initialized using a suboptimal solution produced by a low complexity algorithm, and unsupervised learning-based main training is followed for fine tuning. Through extensive simulations, we confirm that our training methodology with an initialization enables the collection of large amounts of labeled training data within a short preparation time, thereby improving the training performance of the proposed DNN model with a reduced training overhead. Moreover, our results show that the proposed DL-based resource allocation can achieve near-optimal EE, i.e., 95.8% of that of the optimal scheme, with a low computation time of less than 20 milliseconds, which underlines the benefit of the proposed scheme.