LAUSR

Deep learning models for scheduling of potentially-interfering communication pairs, in device-to-device (D2D) settings, require large training samples in the order of hundreds to thousands. Some of the dynamic networks, such as vehicular networks, cannot tolerate the waiting time associated with gathering a large number of training samples. Spatio-temporal correlation among communication pairs in such networks can be utilized to reduce the learning phase. In this paper, we propose a Riemannian-geometric recurrent neural network (R-RNN) method based on statistical recurrent unit (SRU) for wireless link scheduling. First, we represent local graphs around each D2D pair in any finite time frame as a sequence of points on Riemannian manifold thanks to representing its topology as a symmetric positive definite (SPD) matrix. We compute the Riemannian metric, i.e., Stein metric, which are suitable measures of time-dependence among D2D pairs. Then we use the Stein metric in the proposed R-RNN method to forecast the link scheduling decisions for a finite number of successive time slots ahead. Simulation results reveal that the proposed method achieves promising performance against the state-of-the-arts with only 45 training samples.