FULMINA: A Fast Multi-Armed Bandit Approach for Optimal SF Allocation in LoRa IoT Networks

The LoRa (Long Range) protocol is one of the most widely employed technologies in Internet of Things (IoT) networks, thanks to its low-power and long-range communication capabilities. However, dense LoRa networks face challenges in optimizing performance due to limited energy and spectrum resources. While traditional Reinforcement Learning (RL) algorithms represent a valid approach to promote fully-distributed optimization in LoRa networks, they may struggle to adapt to dynamic network conditions and, depending on complexity, may pose a significant computational burden on constrained IoT devices. In light of this, we propose FULMINA (Fast, mULtiarmed bandit approach for optiMal SpreadIng allocation in lorA networks), a novel fully-distributed framework for Spreading Factor allocation in LoRa networks. FULMINA is based on MultiArmed Bandit solutions, a stateless alternative to conventional RL methods. We conducted an extensive simulation campaign, comparing FULMINA with a Q-Learning-based state-of-the-art baseline. The results show that FULMINA is (i) at least 5 times faster than Q-Learning in converging to optimal solutions, and (ii) capable of achieving up to 70 % improvement in average network energy consumption, with a comparable or only slightly reduced network Packet Delivery Ratio.