DEEP REINFORCEMENT LEARNING-BASED EDGE-ASSISTED VIDEO ANALYTICS OFFLOADING

With the development of deep learning, AI-related services and applications have emerged on a large scale, including recommendation systems, video analytics, etc. They all place higher demands on high computing power, high bandwidth, and low latency. Edge computing is now considered to be the most suitable way. In this paper, we consider Multi-user edge-assisted video analytics task offloading (MEVAO) problem, in which all users with different video analytics tasks independently choose their accuracy decisions that satisfy their requirements and then offload the video data to the edge server. In view of this, we propose a deep reinforcement learning approach. The algorithm designed a utility function based on the feature of video analysis, modeled MEVAO as a game theory problem, and achieved the Nash equilibrium. The deep reinforcement learning approach was applied to improve the flexibility of making accurate decisions in different scenarios. Experimental results show that our algorithm has a better performance compared with other existing methods.