Tiny Machine Learning for Efficient Channel Selection in LoRaWAN

Abstract

Machine learning (ML) has emerged as a promising avenue for enhancing the efficiency and intelligence of channel allocation processes. However, deploying ML algorithms on resource-constrained edge devices poses significant challenges due to their limited computational capabilities and storage capacities. In this study, we propose leveraging tiny ML (TinyML) techniques to address these challenges and optimize channel allocation within long range wide area network (LoRaWAN) deployments. Our key innovation lies in replacing traditional random channel allocation methods with TinyML-based approaches, wherein each edge device autonomously utilizes TinyML to select the most efficient channel prior to each uplink transmission. Furthermore, we conduct comprehensive comparisons between TinyML and conventional channel allocation techniques implemented on edge devices. Through extensive simulations, our results demonstrate that TinyML outperforms existing channel allocation mechanisms in terms of packet success ratio (PSR). Notably, when evaluating TinyML against conventional ML approaches in terms of model size and inference time, TinyML exhibits superior performance without compromising efficiency.