边缘计算环境中基于分布式深度强化学习的多域物联网网络任务卸载优化

In the Internet of Things (IoT) networks, sensors, gateways, and services interoperate at different levels to provide services to the end users. IoT networks are deployed in different domains for specific tasks that can be monitored from remote locations. The increase in the number of IoT-connected devices and their notable limited computational power calls for resource-efficient and in-between layers of task processing on the network. In this study, we utilized deep reinforcement learning to intelligently model the offloading policies as Markov Decision Process (MDP) for IoT devices in a distributed manner by considering IoT devices as agents that make offloading decisions taking into account the environmental dynamics. To attain optimal policy in the learning process that caters to high dimensionality, deep Q-network was employed to model the agents’ interaction in a dynamic and environment-sensitive manner. The architecture allows local decision-making by IoT edge nodes for tasks offloading to edge servers based on connectivity, resource availability, and proximity. Extensive simulation under different learning rates, batch sizes, and memory sizes shows that the proposed scheme with the utilization of a CNN approximator generates optimal policy and higher convergence performance with lower latency than the conventional Q-learning model and several other existing algorithms.