Deep Learning-Based Power Control for Uplink Cognitive Radio Networks.

In this paper, we study deep-learning-based power control methods for an underlay cognitive radio (CR) interference channel network, where the SUs are allowed to access the network on the promise of ensuring the quality of service (QoS) of PU. Aiming at boosting the throughput of the whole network, we consider a sum rate maximization power control problem subject to the rate and power constraints of all users. Due to the inter-user interference, the considered problem is nonconvex and thus NP-hard to solve. Different from traditional optimization techniques, we rely on the deep-learning (DL) method to find the solution adaptively. Specifically, we construct a multi-layer fully connected deep neural network (DNN) to deduce the transmit power of PU and SUs thorough self-learning. However, it is not straightforward to apply the classical DNN to solve the sum rate maximization problem. The challenges mainly originate from two aspects. On one hand, the lack of ground truth of the optimal power allocation makes it hard to train the DNN straightforwardly, and on the other hand, the QoS constraints of both PU and SUs in the optimization problem makes things even more complicated. To tackle those difficulties, we adopt unsupervised learning strategy after applying the barrier method to the formulated power control problem. Simulations demonstrate the effectiveness of the proposed DL-based power control method.