Prediction Error Reduction of Neural Networks for Car-Following Using Multi-Step Training

Predicting the surrounding vehicles' behavior is an important requirement for automated driving as it enables estimating others' reactions to the own behavior during planning as well as the identification of critical situations. This work proposes a recursive multi-step training scheme for neural networks that predict other vehicles' positions in a highway car-following scenario. We implement a neural network and compare the proposed approach to the commonly used single-step training as well as parametric models. For this, the Intelligent Driver Model (IDM) and its derivatives have been calibrated using the same approach. Evaluation is performed on 10 hours of real-world car-following situations, extracted from the extensive HighD dataset. Given equal inputs, we show that a minimal neural network with two layers composed of three neurons each surpasses the prediction performance of both the parametric prediction models and the network trained with the standard single-step approach.