Edge intelligence computing for mobile augmented reality with deep reinforcement learning approach

摘要

Convergence of Augmented Reality (AR) and Next Generation Internet-of-Things (NG-IoT) can create new opportunities in many emerging areas, where the real-time data can be visualized on the devices. Integrated NG-IoT network, AR can improve efficiency in many fields such as mobile computing, smart city, intelligent transportation and telemedicine. However, limited by capability of mobile device, the reliability and latency requirements of AR applications is difficult to meet by local processing. To solve this problem, we study a binary offloading scheme for AR edge computing. Based on the proposed model, the parts of AR computing can offload to edge network servers, which is extend the computing capability of mobile AR devices. Moreover, a deep reinforcement learning offloading model is considered to acquire B5G network resource allocation and optimally AR offloading decisions. First, this offloading model does not need to solve combinatorial optimization, which is greatly reduced the computational complexity. Then the wireless channel gains and binary offloading states is modeled as a Markov decision process, and solved by deep reinforcement learning. Numerical results show that our scheme can achieve better performance compared with existing optimization methods.