Transfer Reinforcement Learning Aided Distributed Network Slicing Optimization in Industrial IoT
IR Deposit conditions:
OA version (pathway a): Accepted version
When accepted for publication, set statement to accompany deposit (see policy)
Must link to publisher version with DOI
Publisher copyright and source must be acknowledged
With the growth of the number of Internet of Things (IoT) devices and the emergence of new applications, satisfying distinct QoS in the same physical network becomes more challenging. Recently, with the advance of network functions virtualization and software-defined networking (SDN) technologies, the network slicing technique has emerged as a promising solution. It can divide a physical network into multiple virtual networks, therefore providing different network services. In this article, to meet distinct QoS in industrial IoT, we design a network slicing architecture over the SDN-based long-range wide area network. The SDN controller can dynamically split the network into multiple virtual networks according to different business requirements. On this basis, we proposed a deep deterministic policy gradient (DDPG) based slice optimization algorithm. It enables LoRa gateways to intelligently configure slice parameters (e.g., transmission power and spreading factor) to improve the slice performance in terms of QoS, energy efficiency, and reliability. In addition, to accelerate the training process across multiple LoRa gateways, we leverage the transfer learning framework and design a transfer learning-based multiagent DDPG algorithm. © 2005-2012 IEEE.