Sequence-to-sequence learning for link-scheduling in D2D communication networks

Abstract

Scheduling of Device-to-Device (D2D) links in communication networks conventionally relies on solving NP-hard combinatorial optimization problems. These types of solution approaches will not be suitable for the service requirements of future networks due to the associated computational complexity. That is why Deep Learning (DL) is one of the promising approaches to tackle this problem. Nevertheless, designing the learning algorithm to cope with the dynamic nature of the D2D network is a challenge. Current research using DL only assumes a static layout of the network without taking advantage of the correlations between the decisions in a dynamic network. Consequently, this paper proposes a sequence-to-sequence modelling (SSM) method for D2D scheduling using only distance information. The SSM processes the distance information as well as the previous scheduling decisions in a sequential manner with a feedback from the intermediate output, and models the correlations between consecutive input information as well as the produced decisions. Simulation results show that the average sum rate of the SSM is about 95% of that achieved by the optimal scheduler and it requires at least 90% less resources than those required by other DL schedulers reported in the literature. Finally, the decision-making of SSM is explored for key input descriptors and an unsupervised decision-maker is explored, which is shown to produce reasonable results with minimal computational requirements.