Resource Allocation of IoT systems Integrated with Blockchain and Mobile Edge Computing

• Published in: 2022 4th International Conference on Frontiers Technology of Information and Computer (ICFTIC) pp. 377 - 382
• Main Authors: Bai, Zihan, Wan, Jianxiong, Li, Leixiao, Liu, Chuyi, Duan, Mingda
• Format: Conference Proceeding
• Language: English
• Published: IEEE 02.12.2022
• Subjects: Blockchain; Blockchains; Deep Reinforcement Learning (DRL); Energy consumption; Internet of Things; Internet of Things (IoT); Mobile Edge Computing (MEC); Multi-access edge computing; Reinforcement learning; resource allocation; Scalability; Simulation
• DOI: 10.1109/ICFTIC57696.2022.10075287

With the development of the Internet of Things (IoT), IoT devices have been widely applied into several fields to collect and transmit data. However, it is crucial to ensure to the security of the collected data. On the other hand, computing resources of IoT devices have obvious constraint. Blockchain and Mobile Edge Computing (MEC) can significantly improved the security of the IoT system and the efficiency of the consensus process of IoT devices. However, they also bring a lot of energy consumption and computational overhead. Reasonable allocation of computational resources is an effective method to reduce the energy consumption and the computational overhead. The resource allocation problem of the IoT system supported by blockchain and MEC can be formulated as an Markov Decision Process (MDP). In the existing studies, Deep Q-Network (DQN)-based approaches are adopted to optimize the energy consumption and the computational overhead. However, as the dimensions of the actions become larger, the action space of DQN-based approaches will have scalability limitations. Therefore, we propose the Branching Dueling Q-Network Resource Allocation (BDQ-RA) algorithm to address the problem of scalability limitations. In this article, we consider the profit between the earnings of computational tasks and the weight cost as the reward. Simulation results show that our algorithm can reduce the action space to one ninth and improve the reward about 12% compared with DQN solutions.