Federated Reinforcement Learning for Automated LoRaWAN Management in Industrial IoT

The scale of wireless networks is growing rapidly to support next generation industry 5.0 applications. LoRaWAN offers features such as low power consumption and long communication range, beneficial in many industrial applications. However, this comes at the cost of low data rates and link margins that rely on coding gain and careful network planning. These qualities make LoRaWAN networks vulnerable to interference and data-loss, which are particularly problematic in large-scale and mobile networks. It is vitally important that LoRaWAN networks are automatically and adaptively managed, to ensure that resources are optimally allocated, while satisfying multiple application requirements. In this paper, a novel multi-agent federated reinforcement learning framework is designed to enable knowledge sharing across multiple agents, and to resolve the issue of limited sampling coverage deep reinforcement learning. The goal is to allow individual node agents to adaptively control their network parameters to satisfy their own application requirements, while ensuring fair access across all nodes. The proposed method is evaluated against a non-federated approach. The results show that the proposed federated reinforcement learning approach can overcome the cold-start problem when new devices join a network. The proposed approach also demonstrates the ability to support mobile nodes travelling through locations, with changing network conditions, while maintaining communication performance.