A QoS-Aware Technique for Computation Offloading in IoT-Edge Platforms Using a Convolutional Neural Network and Markov Decision Process

• Published in: IT professional 2023-01, Vol.25 (1), p.24-39
• Main Authors: Heidari, Arash, Jamali, Mohammad Ali Jabraeil, Navimipour, Nima Jafari, Akbarpour, Shahin
• Format: Article
• Language: English
• Subjects: Artificial neural networks; Batteries; Computation offloading; Decision making; Deep learning; Delays; Energy consumption; Energy harvesting; Failure rates; Internet of Things; Markov analysis; Markov processes; Modules; Network latency; Neural networks; Quality of service; Reinforcement learning; Transfer learning
• ISSN: 1520-9202; 1941-045X
• DOI: 10.1109/MITP.2022.3217886

Offloading is one of the critical enablers of the Internet of Things (IoT) as it helps overcome the resource limitations of individual objects. Offering enough computational power for IoT applications at the edge has become a severe problem. An intelligent edge is a potential approach for pushing intelligence to network edges, which has played the role of intelligent decision-making in many elements of edge, notably task offloading. By leveraging the edge, IoT devices with limited battery capacity can offload a portion of the tasks that can dramatically reduce latency and improve battery life. Because IoT devices have limited battery capacity, employing deep learning approaches in such devices results in higher energy consumption. As a result, several studies used energy harvester modules, which are not available to IoT devices in real-world scenarios because many IoT devices lack such modules. This article proposes the offloading problem by leveraging the Markov decision process. Furthermore, we built a lightweight version of the reinforcement learning technique to decrease complexity and deployed it in IoT devices. Then, we used a convolutional neural network to accelerate learning and put it on the edge platform. Throughout the entire working duration of the system, these two methods collaborate to provide the optimal offloading strategy. Also, transfer learning was used to initialize Q-table values to improve the system's efficiency. The results showed that the proposed method outperforms five benchmarks in terms of delay by 3.3%, IoT device efficiency by 3.2%, energy use by 4.2%, and task failure rate by 2.9% on average.