Impact of Multi-Layer Recurrent Neural Networks in the Congestion Analysis of TeraHertz B5G/6G MAC Mechanism

Abstract

Nowadays design of BSG/6G radio technologies require analysis based on simulations to determine optimum functioning properties. We executed ns-3 simulations to generate TeraHertz scale MAC event sequences. Standard communication proposal mechanism, called Adaptive Directional Antenna Protocol for Terahertz (ADAPT), was analysed by extract frame collision behaviour in the control plane of the high-speed channel. Seven step sizes of sector indexes with specific features were used at the base station to give access to the mobile terminals spread in 30 sectors of the circular radio cell. After presenting basic properties of the MAC mechanism we grouped collision sequences into four classes. Testing classifications were performed with three types of recurrent neural networks (RNN). Transfer learning was used to detect influence of the recurrent layers on the performance of the compound multilayer RNN. Complex metric was introduced to quantify the learning efficiency of the RNN. It was found that the proposed metric, called Weighted Accuracy-to- Time Ratio is able to characterize and compare in efficient manner goodness of different deep learning techniques used for evaluation of the ADAPT technology. This new metric quantifies transfer learning property and differentiates applicability of the most popular recurrent neural networks used in practice.