ALLEVIATE COLLISIONS IN LORA NETWORKS USING REINFORCEMENT LEARNING

Abstract

As Low Power Wide Area Networks (LPWANs) continue to expand to support the increasing demands of Internet of Things (IoT) applications, they face major limitations in terms of scalability, collision management, and network reliability. These challenges are particularly pronounced in LoRaWAN, a widely adopted LPWAN protocol that relies on ALOHA-based medium access mechanisms. As network density increases, lack of coordination in ALOHAbased transmission leads to high collision rates and decreased packet delivery performance. In this work, we propose a novel reinforcement learning (RL)-driven framework that enhances LoRaWAN performance by introducing intelligence at the edge, without requiring changes to the existing protocol stack. Our solution leverages the SARSA algorithm to enable enddevices (EDs) to autonomously learn optimal transmission slots based on their local experience. A lightweight synchronization scheme ensures that slot selection remains consistent across devices, while preserving LoRaWAN compatibility. To optimize learning behavior, we perform comprehensive hyperparameter tuning and evaluate policy generalization through transfer learning experiments. The entire framework is deployed and tested in a real-world testbed built using MicroPython on ESP32-S3, and custom network server. Experimental results show that our RL-based approach achieves over 36% improvement in Packet Delivery Ratio (PDR) compared to traditional Pure ALOHA and Slotted ALOHA methods, with only minimal energy overhead. To promote reproducibility and support future innovation in this area, we provide open-source implementations of the testbed and protocol logic.