Adaptive learning in low-power wide-area networks

Internet of Things applications are driving the need for better and advanced solutions to connect and manage sensors communicating into a smarter world. In this thesis we present preliminary results derived from tests on dual-band Low Power Wide Area Networks, trying to solve both effectiveness and efficiency challenges from a data science point of view. In the former case, by applying machine learning models with high performances while keeping the load of sensor networks as low as possible in terms of acknowledgements requested by single nodes, in the latter by storing a simple model in a small device while providing fast and accurate predictions. Statistical inference techniques and online learning algorithms are investigated under non-stationary conditions and adopted for testing multiple practical scenarios, with the aim of estimating the network status in the licensed and unlicensed bands, i.e. NB-IOT and Lorawan respectively.