PERFORMANCE ENHANCEMENTS OF LORAWAN USING MACHINE LEARNING ON THE EDGE

Abstract

LoRaWAN is becoming the dominant long range protocol for Internet of Things (IoT) devices. However, LoRaWAN’s performance suffers from a high number of collisions in saturated LoRa networks. To mitigate the number of collisions that happen due to the time overlap of transmissions on the same channel, we use an edge machine learning approach. To do so, Reinforcement learning (RL) is leveraged. RL is a field of machine learning that aims at maximizing the reward by interacting with the environment. SARSA is an on-policy RL algorithm that uses previous actions to update the Q-value. This study aims to improve the performance of congested LoRa networks by allowing RL-based applications on individual nodes. Specifically, it explores whether periodic applications driven by SARSA can improve the performance of the network and adapt the period transmissions of the nodes. In this thesis, two versions of SARSA have been developed, evaluated, and compared to the baseline of LoRaWAN. To achieve that, several simulations with different configurations are performed. The simulations include networks with hundreds of nodes and different number of maximum retransmissions. The results of the simulations have shown that networks where SARSA algorithms are used present a better performance compared to the typical LoRaWAN periodic application in certain examined scenarios. The results demonstrate that RL-based algorithms can significantly improve the performance of networks with high load. Nevertheless, there is still room for further improvement and better understanding of the internal mechanisms of the proposed RL approaches.