REINDEAR: REINforcement learning agent for Dynamic system control in Edge-Assisted Augmented Reality service

Abstract

Nowadays the industry of the Augmented/Virtual Reality is rapidly growing in various domains such as AI assistance or gaming. While the main goal of the AR service itself is to provide the user the visualized computing experience in their life, the systematic requirements include the real-time performance since the service should be reacting to the user's state which changes dynamically over time. This kind of real-time response is often very hard to achieve on typical edge devices such as mobile phone or AR goggles due to lack of computational power. To overcome this issue, many of the researches suggest server offloading technique which can provide sufficient amount of computational power in exchange of the transmission overhead. The tradeoff relationship between the computational power and transmission overhead makes the control of the offloading procedure important for the overall service quality. In this paper we propose an RL based server-client controlling scheme REINDEAR. REINDEAR is a system that conducts class-wise characteristic analysis from the experience, so that it could control the AR service tradeoff quality adaptively. From the result of the experiments, we showed that our REINDEAR system learns the underlying behavioural patterns of video objects and provides controls that suits each pattern.