Dynamic Parameter Allocation With Reinforcement Learning for LoRaWAN

Abstract

LoRaWAN attracted lots of attention with its capacity for large device numbers, long-range and low power consumption. In order to simplify the transmission procedure, a pure Aloha protocol is implemented into its MAC layer. However, as the number of connected devices to the base station increases, the devices’ transmission parameters allocation becomes a vital issue related to network performance. This research contributes to the decentralized dynamic Spreading Factor (SF) allocation strategies during transmission by proposing STEPS, a Score Table based Evaluation and Parameters Surfing approach. STEPS is a reinforcement learning-based method that evaluates and changes the parameters based on probability and score tables. It provides a nondeterministic parameter selection method by updating the table while transmitting. Some variants of STEPS with different algorithms are proposed. Moreover, an estimation-based initialization is proposed to improve learning performance. Simulations and statistical tests are carried out with MULANE, a lightweight LoRaWAN Simulator developed in our previous work. The results show that the estimation has a high confidence level. Compared with the baseline methods, the proposed methods reduce energy consumption by 24-27% in different numbers of nodes. For bi-directional transmission, the proposed methods increase the 18% network throughput in a small number of nodes and 33% in a large number of nodes. Moreover, the proposed methods provide a framework of decentralized parameter allocation, which gives the extendability of this work.