Client-Based Intelligence for Resource Efficient Vehicular Big Data Transfer in Future 6G Network
Vehicular big data is anticipated to become the "new oil" of the automotive industry which fuels the development of novel crowdsensing-enabled services. However, the tremendous amount of transmitted vehicular sensor data represents a massive challenge for the cellular network. A promising method for achieving relief which allows to utilize the existing network resources in a more efficient way is the utilization of intelligence on the end-edge-cloud devices. Through machine learning-based identification and exploitation of highly resource efficient data transmission opportunities, the client devices are able to participate in overall network resource optimization process. In this work, we present a novel client-based opportunistic data transmission method for delay-tolerant applications which is based on a hybrid machine learning approach: Supervised learning is applied to forecast the currently achievable data rate which serves as the metric for the reinforcement learning-based data transfer scheduling process. In addition, unsupervised learning is applied to uncover geospatially-dependent uncertainties within the prediction model. In a comprehensive real world evaluation in the public cellular networks of three German Mobile Network Operators (MNOs), we show that the average data rate can be improved by up to 223 % while simultaneously reducing the amount of occupied network resources by up to 89 %. As a side-effect of preferring more robust network conditions for the data transfer, the transmission-related power consumption is reduced by up to 73 %. The price to pay is an increased Age of Information (AoI) of the sensor data.