Reinforcement Learning Based Approach for Virtualized Face Detection at the Edge

Real-time requirements in video streaming and processing are increasing and represent one of the major issues in industry 4.0 domains. In particular, Face Detection (FD) use-case has attracted the interest of industrial and academia researchers for various applications such as cyber-physical security, fault detection, predictive maintenance, etc. To ensure applications with real time performance, Edge Computing is a good approach which consists in bringing resources and intelligence closer to connected devices and hence, it can be used to cope with strong latency and throughput expectations. In this paper, we consider optimal routing, placement and scaling of virtualized face detection services at the edge. We propose an edge networking approach based on Integer Linear formulation to cope with small problem instances. A reinforcement learning solution is proposed to address larger problem sizes and scalability issues. We assess the performance of our proposed approaches through simulations and show advantages of the reinforcement learning approach to converge towards near-optimal solutions in negligible time.