Predicting Driver Adaptability: A Hierarchical Computational Rationality Model of Bounded-Optimal Visual Sampling in Lateral and Longitudinal Vehicle Control

Abstract Attention allocation plays a central role in how drivers adapt to uncertainty resulting from human attentional and perceptual limitations, particularly when visual information must be sampled intermittently while maintaining control of the vehicle. However, how drivers adapt their visual sampling behavior when uncertainty arises simultaneously in several part-tasks of driving, such as lateral and longitudinal vehicle control, remains insufficiently understood. We developed a computational rationality model to predict drivers’ adaptation of attention allocation in a driving task requiring lane keeping and speed maintenance under visual occlusion. Using reinforcement learning in a simulated environment, the model employs separate but interacting policies for lateral control, longitudinal control, and supervisory visual sampling of task-relevant information. Attention-dependent internal state estimates evolve as a function of intermittent sensory access, control actions and ongoing vehicle motion and can increasingly diverge from the external environment. In contrast to earlier work, the present model extends the lateral control task to include continuous speed regulation and evaluates the integrated model on a curved road with sinusoidal elevation changes. The model showed good aggregate correspondence with human data collected from 30 participants in a driving simulator, with all eight evaluated model means falling within one human standard deviation of the corresponding human means. The model provides a formal account of how adaptive visual sampling may emerge when lane keeping, speed regulation and limited visual information must be balanced under uncertainty. More broadly, the framework supports extension to additional driving subtasks and future investigation of inter-individual differences in attentional adaptation.