On Latency Prediction with Deep Learning and Passive Probing at High Mobility

In autonomous driving, several applications like teleoperated driving, back-end status verification, or online gaming for customer infotainment rely on low-latency communication. Ideally, we can select a route that best supports the applications’ requirements before the journey. Therefore, route selection for autonomous vehicles might require in-advance latency predictions. End-to-end (E2E) latency prediction is a difficult task, especially when considering that it needs to be achieved with limited active probing due to cost constraints. We study continuous latency prediction and application feasibility assessment (in terms of meeting the applications’ E2E latency requirements), using a custom-designed deep learning model that leverages feature engineering for prediction error reduction. We provide insights into the model behavior utilizing recent advances in explainable artificial intelligence. Moreover, we present a novel model-agnostic approach based on active learning to leverage passive probing data. A pre-trained model performs certainty sampling, predicts artificial labels to enlarge the training dataset, and trains iteratively on the augmented set. The results show a 5 % reduction in mean average error for continuous latency prediction and an increase of up to 2.8 % in macro F1 score due to the use of passive probing data.