Enhanced cellular bandwidth prediction for highly automated driving

Abstract

Highly automated vehicles will change the future of our personal mobility. To ensure safety and comfort while driving its passengers, the vehicle has to rely on the newest traffic and map updates at any time. Furthermore the passengers want to enjoy infotainment services like video and music streaming during the travel. All these kinds of services require a reliable and fast cellular internet connection. However due to the high speed of the vehicle and the varying deployed cellular infrastructure, the experienced network throughput is constantly changing. To predict those throughput changes and to maintain the overall experienced network quality at a high level, machine learning techniques can be leveraged. In our previous work [1], we first investigated the idea to train machine learning techniques based on specifically localized training data provided by modern day smartphone APIs in a so-called connectivity map to improve their overall performance. From the first promising results obtained in this work, we now further improved the quality of our input feature set in this current work by introducing more precise lower level protocol information to the prediction process. The measurements were obtained from the same real highway driving scenario over a period of three days, in which over 540.000 precise lower layer throughput estimations could be collected. Based on this more accurate data set, we were able to improve the overall prediction accuracy and clearly showcase the performance gains achieved through localized training data in comparison to a general global training data set.