Attention Allocation in a Cognitive Driver Model (AIE / CASCaS)

emerging Updated: 2026-06-08
attention-allocation visual-attention cognitive-architecture driver-model gaze-behaviour SEEV multitasking situation-awareness simulation methodology

Attention Allocation in a Cognitive Driver Model (AIE / CASCaS)

Status: emerging
Last updated: 2026-06-08
Sources: Wortelen_Luedtke_Baumann2013 Integrated_Simulation_Of_Attetnion_Distribution_And_Driving_Behavior.Pdf
Tags: [attention-allocation, visual-attention, cognitive-architecture, driver-model, gaze-behaviour, SEEV, multitasking, situation-awareness, simulation, methodology]

Summary

Wortelen, Lüdtke and Baumann (2013) present a computational driver model that simulates how visual attention is distributed across competing tasks and how that distribution shapes driving performance. The model uses the Adaptive Information Expectancy (AIE) model of attention allocation, embedded in the CASCaS cognitive architecture, to assign each task an attentional weight from its information-event rate and its priority, then select tasks probabilistically. Compared against driving-simulator data, the model reproduced the main effects of event rate and task value on gaze behaviour — percentage dwell time and glance frequency — though with smaller magnitude, and it failed to capture drivers' adjustment of glance duration under load. The work is a model of the same gaze behaviour this knowledge base measures empirically, approached from the generative side.

Body

Context

Wortelen, Lüdtke and Baumann (2013), at a BRIMS conference and within the EU D3CoS project, build and test a cognitive driver model whose purpose is to predict visual attention distribution in a multitasking driving setting and to evaluate the AIE model of attention allocation (PDF p. 2). Their method pairs simulation with a driving-simulator study of human drivers, comparing model and human gaze behaviour and task performance. Within this knowledge base the article sits in the spatial-attention and attention-allocation strand: it is the generative counterpart to the empirical driving work in Visual Occlusion Attentional Demand, and its model-versus-data comparison is an applied instance of the modelling standards in Cognitive Model Validation. Because it predicts dwell times and glance frequencies, it connects directly to gaze-metric measurement in Fixation Saccade Detection.

Key Points

Attention is modelled as goal-directed and resource-limited. Wortelen et al. adopt the eye-mind assumption (Just & Carpenter, 1980) that visual and mental attention are tightly coupled, so a model that allocates mental attention to tasks also predicts where the eyes go (PDF p. 3). Mental attention is defined by the currently active goal, and the AIE model applies where several tasks must be performed in a time-shared way — the same situation as a pilot monitoring displays while controlling an aircraft. Tasks are represented as hierarchical goal models in GOMS-style IF–THEN rules (Card, Moran & Newell, 1983) interpreted by CASCaS (PDF p. 3).

The AIE model operationalises attention from information-event rates. Where the SEEV model is used to predict scanning across information sources, the AIE model predicts which task the agent attends to (PDF p. 3). It assigns each active goal an attentional weight that combines the task's value (a rank-ordering set by the modeller) and its expectancy — how much new information the agent expects for that task — with the next goal selected probabilistically from the relative weights (PDF pp. 3–4). The distinctive contribution is automating the expectancy term: CASCaS counts "information events" (a percept rule directing gaze to a source, followed by a regular rule using that information) and derives event rates from the task model's interaction with the environment, rather than having the modeller set them by hand (PDF p. 4). Wortelen et al. note the unresolved question of whether value and expectancy should combine additively or multiplicatively, and use the additive form, which earlier work found fit better (PDF pp. 3–4).

The driver model reproduced attention effects but underestimated them. The model performed three concurrent tasks — lateral (lane-keeping), longitudinal (speed-keeping near 100 km/h), and a secondary in-vehicle number-reading task (PDF pp. 4–5). Against simulator data, rising event rates or task values raised the percentage of time drivers looked at the supporting information source, and the model reproduced these effects on dwell time and, in most cases, on glance frequency — but the magnitude was usually smaller for the model (PDF pp. 6–7). Task-priority manipulation revealed a clear discrepancy: human task performance was best for whichever task was prioritised, whereas in the model prioritising the lateral control task improved all three tasks, giving a poor correlation with human reaction times despite small absolute deviations (PDF p. 7).

The model lacks a duration-of-attention mechanism. Wortelen et al. trace the discrepancy to a missing mechanism: drivers shortened their average glances to the secondary display and the speedometer as road curvature rose, even though those tasks were unchanged, suggesting attention adapts in duration as well as frequency under pressure (PDF pp. 7–8). CASCaS executes a task in a fixed amount of time aside from noise, so it cannot shorten or lengthen task processing with priority; the authors propose that rising priority should raise not only selection probability but also task-execution accuracy and the duration of stabilisation phases (PDF p. 8).

Conclusion

Wortelen et al. (2013) conclude that the cognitive driver model, built to study visual attention distribution, also inherently reproduces effects on task performance — especially sensitivity to the event rate of the lateral control task — but fails to capture some human effects because manipulating event rates and priorities changes not just the frequency of attention to a task but its duration, which the model does not represent (PDF p. 8). They flag the duration mechanism, the additive-versus-multiplicative combination of value and expectancy, and the learning/convergence of event functions as future work. For this knowledge base the value is a concrete, falsifiable model of gaze allocation: it predicts the dwell-time and glance-frequency measures the empirical articles record, and its documented failure mode (glance duration under load) is itself a hypothesis about gaze behaviour worth testing.

References

Card, S.K., Moran, T.P. and Newell, A. (1983) The Psychology of Human-Computer Interaction. Hillsdale, NJ: Erlbaum. To be validated.

Just, M.A. and Carpenter, P.A. (1980) 'A theory of reading: From eye fixations to comprehension', Psychological Review, 87(4), pp. 329–354. To be validated.

Salvucci, D.D. (2006) 'Modeling driver behavior in a cognitive architecture', Human Factors, 48(2), pp. 362–380. To be validated.

Wickens, C.D., Helleberg, J., Goh, J., Xu, X. and Horrey, W.J. (2001) Pilot Task Management: Testing an Attentional Expected Value Model of Visual Scanning. Technical Report ARL-01-14/NASA-01-7. Moffett Field, CA: NASA Ames Research Center. To be validated.

Wortelen, B., Lüdtke, A. and Baumann, M. (2013) 'Integrated Simulation of Attention Distribution and Driving Behavior', in Proceedings of the 22nd Annual Conference on Behavior Representation in Modeling and Simulation (BRIMS 2013). wortelen2013integrated

Open Questions

  • The model uses the eye-mind assumption to treat task attention as a proxy for gaze. How well does that hold for the KB's VR/HMD gaze data, where covert attention and overt gaze can decouple?
  • The documented failure — drivers shorten glance durations under load while the model holds task time fixed — predicts a measurable gaze signature (curvature-dependent glance-duration shortening). Is this reproduced in other gaze datasets in the corpus?
  • AIE derives expectancy from a formal task model. Could empirically measured event rates from eye-tracking data calibrate or replace the modelled expectancy term?