Predicting driver adaptability: a computational rationality model for attention allocation in lateral vehicle control
该研究构建了一个计算理性模型，结合强化学习预测驾驶员在横向车辆控制中如何适应不确定性来分配视觉注意力，模型能复现人类行为模式，有助于设计驾驶员注意力监测和辅助系统。
Attention plays a pivotal role in understanding human adaptability under uncertainty, particularly in directing perceptual information acquisition. However, how individuals adapt visual sampling to dynamic uncertainties in internal task-state representations remains poorly understood. We developed a computational rationality model to predict drivers’ adaptation to uncertainty in lateral vehicle control. Using reinforcement learning within a simulated task environment, we optimised attention allocation decisions to maximise lane-keeping performance under partial observability by comparing model predictions with human data from two simulator studies. The task required drivers to maintain lane position during visual occlusion as long as possible, with the model demonstrating promise for understanding how drivers adapt visual sampling to dynamic and subjective uncertainties. The model reproduced key human-like behavioural patterns and provides a rational basis for estimating how drivers decide to sample visual information. Computational rationality seems to enable simulation and prediction of drivers’ sampling while allowing adjustment for personal parameters.Practitioner summary Using computational rationality with reinforcement learning, the model predicts when drivers sample visual information and how they steer under occlusion. Adjustable parameters (steering skill and internal cautiousness) enable individualised fits capturing inter-driver variability. This informs early-stage driver-attention monitoring and assistance design, strengthening safety systems and reducing reliance on laboratory testing.