Modeling an intelligent controller for predictive caching in AR/VR-enabled home scenarios

Abstract

In a possible future, pervasive augmented and/or virtual reality (AR/VR) might become the primary delivery method for audio-visual information. To achieve a high level of user satisfaction, such a system must answer user requests with high-quality content delivered with minimal lag. As bandwidth and latency limitations will still apply, the system must perform predictive caching of the content. In this paper, we investigate several strategies for predicting the information needs of a user in an AR/VR-enabled home. The paucity of datasets is a major challenge in such studies. We are starting from the hypothesis that the user’s patterns of daily life guide the content consumption regardless of the delivery medium. This allows us to synthetically generate realistic content requests starting from real-world databases of user activities in smart homes. Using these datasets, we develop techniques for demand prediction and content caching that aim to optimize the quality of user satisfaction while minimizing the cost of caching. We propose three algorithms: one based on probabilistic modeling, one based on long short term memory (LSTM) networks, and one based on majority voting. Through a set of experimental studies, we show that our techniques outperform baseline caching techniques both in terms of user satisfaction and caching cost.