Exploiting Future Radio Resources With End-to-End Prediction by Deep Learning

Machine learning is a powerful tool to predict user behavior and harness the vast amount of data measured in cellular networks. Predictive resource allocation is a promising approach to take advantage of the prediction for the mobility and traffic load related user behavior. This paper strives to boost the performance of under-utilized networks by predicting behavior-related information from historical data with deep learning. We first propose a hierarchical and multi-timescale radio resource management scheme for non-realtime service that only needs coarse-grained future knowledge, by taking multi-input-multi-output orthogonal frequency multi-access as an example system and high throughput as an example performance metric. Such a scheme allows the decision of resource management to be made in a central processor and base stations in different timescales and allows the knowledge to be predicted with less training samples. Then, we design a deep neural network to learn the future knowledge required for making decision directly from different types of past data with different resolutions observable in cellular networks. Simulation results show that the proposed scheme with the end-to-end knowledge prediction performs closely to the relevant optimal solution with perfect and fine-grained prediction, and provides dramatic gain over non-predictive counterpart in supporting high request arrival rate for the non-realtime service.