ACT: Optimizing Lorawan via Comprehensive Modeling and a Deep Reinforcement Learning Resource Allocation Strategy

Long Range Wide Area Network (LoRaWAN), a Low Power Wide Area Network (LPWAN) technology based on Long Range (LoRa), is extensively studied and applied but struggles with bandwidth limitations as device numbers increase. This necessitates optimized resource allocation strategies to enhance network performance. Current methods typically focus solely on the parameters of the uplink, such as the uplink spreading factor and transmission power, while neglecting the settings of the downlink, which include the downlink spreading factor and the size of the receive window. The configuration of the downlink directly determines whether devices can correctly receive the parameters distributed by the gateway. This neglect leads to an incomplete understanding of the performance of LoRaWAN networks. This paper introduces an Actor-Critic algorithm based on global Transformer (ACT) to address these challenges. ACT utilizes online Deep Reinforcement Learning (DRL) to continuously optimize uplink and downlink parameters based on real-time interactions. To evaluate the effectiveness of ACT, we enhanced an existing LoRa/LoRaWAN simulator to create an open-source LoRaWAN simulation tool named LoRaWANSim. LoRaWANSim accounts for the demodulator lock effect and supports multi-channel, multi-demodulator gateways for bidirectional communication, both of which are ignored by most other simulators. Comparative experiments across various device densities and traffic intensities show that ACT improves Packet Delivery Rate (PDR) and Energy Efficiency Ratio (EER) by 10 % and 39 %, respectively.