On the Detection of Spectrum Irregularities through Deep Learning in Dense IoT architectures

In this work, a novel Deep Learning model that combines Q Learning and Neural Networks is proposed and experimentally evaluated. The proposed scheme is developed in order to detect and tackle the effects of spectrum anomalies, which are unexpectedly appeared in ultra-dense Internet of Things (IoT) architectures. The protocol which is taken into consideration in this work, is the IEEE 802.11ah (Wi-Fi HaLow) and it is tested under strong interference, caused between wireless links which transmit in non-overlapping frequencies. The proposed approach trains a model that constantly observes the wireless environment and obtains an optimal policy for the transmission time parameter re-configurations of the participating devices. The experimental evaluation showcases that the proper training of the proposed Deep Q Learning model, leads to remarkable increased Packet Delivery Ratio (PDR) and throughput in the examined scenarios. Apart from the improvements (+35%) observed on a PDR and throughput basis, the proposed algorithm also achieves overall higher channel utilization, increased transmission opportunities, and fairness in terms of channel access.