An Adaptive Neural Network Characterization of Driver Longitudinal Control Behavior

Abstract

This paper examines how neural network methods may be used to represent driver manual control of throttle position during headway-keeping tasks. The findings of the study indicate that the neural network methodology employed here can be used to characterize a variety of driver longitudinal control behaviors, provided that the input data exhibit strong similarities to the data used to train the network. Otherwise, the network architecture and methodology utilized here is not considered adequate to predict driver control responses for unseen data not represented by the training set. Shortcomings in extending the trained network to predict accurate driver control responses for unseen data can stem from a variety of reasons, including the intermittent nature of driver longitudinal control behavior. While neural networks can be used to represent adaptive longitudinal control behavior of drivers and their associated variability during active control engagements, determining when active driver control engagements are actually occurring and selecting that data for training purposes is a confounding factor.