Deep Q-learning Network Solution for Subjective Computational Task Scheduling in Edge Platforms

The core concepts of edge intelligence have intensified and are permeating into edge computing analytical frameworks, with practical application in 6G and other beyond 5G (B5G) networks. One important area of application of edge intelligence in advancing 6G and other B5G networks is in video analytics. Given the increasing important role of video analysis for security purposes such as in real-time monitoring and dynamic situational adaptation, the incorporation of edge video analytics is considered as an indispensable element for characterizing intelligent task offloading in edge platforms. For this to work, edge storage infrastructure and computational capacity must be significantly improved to ensure optimal performance. To achieve this, the artificial intelligence (AI)-based edge systems and solutions being employed must be able to learn to distinguish between data and video traffic, which is the focus of this paper. To improve edge video analysis, therefore, the paper proposes a computational task scheduling scheme for edge analytics that uses congestion as a subjective metric to measure the device-centric risk-based satisfaction of the offloading devices. Then, a deep Q-learning network (DQN) architecture is used to facilitate transfer learning processes among smart devices in a cooperative environment. The results show that the proposed algorithm achieves better satisfaction and offloading ratios by reducing congestion. Also, the energy consumption achieved in the video analysis using the DQN-based cooperative learning is shown to be 19% less than the traditional individual learning that uses the traditional deep reinforcement learning (DRL) strategy.