A hybrid approach for latency and battery lifetime optimization in IoT devices through offloading and CNN learning

• Published in: Sustainable computing informatics and systems Vol. 39; p. 100899
• Main Authors: Heidari, Arash, Navimipour, Nima Jafari, Jamali, Mohammad Ali Jabraeil, Akbarpour, Shahin
• Format: Journal Article
• Language: English
• Published: Elsevier Inc 01.09.2023
• Subjects: Convolutional neural network; Deep reinforcement learning; Edge; Energy utilization; IoT; Markov decision process; Offloading; Energy utilization; Convolutional neural network; Deep reinforcement learning; Markov decision process; Offloading; Edge; IoT
• ISSN: 2210-5379
• DOI: 10.1016/j.suscom.2023.100899

Offloading assists in overcoming the resource constraints of specific elements, making it one of the primary technical enablers of the Internet of Things (IoT). IoT devices with low battery capacities can use the edge to offload some of the operations, which can significantly reduce latency and lengthen battery lifetime. Due to their restricted battery capacity, deep learning (DL) techniques are more energy-intensive to utilize in IoT devices. Because many IoT devices lack such modules, numerous research employed energy harvester modules that are not available to IoT devices in real-world circumstances. Using the Markov Decision Process (MDP), we describe the offloading problem in this study. Next, to facilitate partial offloading in IoT devices, we develop a Deep Reinforcement learning (DRL) method that can efficiently learn the policy by adjusting to network dynamics. Convolutional Neural Network (CNN) is then offered and implemented on Mobile Edge Computing (MEC) devices to expedite learning. These two techniques operate together to offer the proper offloading approach throughout the length of the system's operation. Moreover, transfer learning was employed to initialize the Q-table values, which increased the system's effectiveness. The simulation in this article, which employed Cooja and TensorFlow, revealed that the strategy outperformed five benchmarks in terms of latency by 4.1%, IoT device efficiency by 2.9%, energy utilization by 3.6%, and job failure rate by 2.6% on average. •Establishing a structure for dynamic offloading for different IoT edge applications;•Creating the MDP model and introducing the QIEO approach to describe the issue of IoT-edge computing offloading;•Reducing data storage and computation by representing state-action interaction in the Q-learning process•Improving system performance by reducing energy consumption, computing delay, device efficiency, and job failure rates.