Expected-Value Model of Visual Scanning (SEEV) in Multitask Flight

emerging Updated: 2026-06-13
visual-scanning selective-attention attention-allocation SEEV expected-value-model multiple-resources percentage-dwell-time divided-attention aviation situation-awareness eye-tracking

Expected-Value Model of Visual Scanning (SEEV) in Multitask Flight

Status: emerging
Last updated: 2026-06-13
Sources: Hfes.45.3.360.27250.Pdf
Tags: [visual-scanning, selective-attention, attention-allocation, SEEV, expected-value-model, multiple-resources, percentage-dwell-time, divided-attention, aviation, situation-awareness, eye-tracking]

Summary

Wickens, Goh, Helleberg, Horrey and Talleur (2003) test two models of how a pilot divides attention across competing tasks and information sources while flying a high-fidelity simulator with a head-mounted eye tracker. The first part asks whether moving traffic and communications information from the ear (ATC voice) to the eye (cockpit displays) helps or hurts; the data show a modest cost of visual delivery, mediated by head-down scanning, consistent with a multiple-resource account rather than auditory "preemption", and no benefit from presenting information redundantly. The second part fits an optimal expected-value model of visual scanning — the expectancy-and-value core of the SEEV model — to the proportion of time pilots dwell in each area of interest. The model reproduces scanning with high accuracy across three experiments, establishing a computational, gaze-based standard for how selective visual attention is allocated under task priority.

Body

Context

Wickens et al. (2003) report a two-part flight-simulation study (12 instrument-rated pilots in a Frasca 142 simulator, eye and head movements sampled at 60 Hz with an ASL Model 501 head-mounted tracker accurate to better than 1°) (PDF p. 5, orig. p. 363). Part 1 manipulates the modality of two side tasks — traffic information (a cockpit display of traffic information, CDTI, vs. ATC voice) and communications (a data link text display vs. ATC voice) — against the higher-priority visual tasks of aviating and navigating, and reads the results through multiple-resource and preemption theories of divided attention. Part 2 takes the eye-tracking record from the same task and fits a computational model of selective attention. Within this knowledge base the article is the aviation anchor of the attention-allocation strand: it is the published expected-value scanning model that the cognitive driver model in Attention Allocation Driver Model builds on (the AIE model extends this SEEV line), its model-against-data fitting is an applied case of Cognitive Model Validation, and it shares the limited-resource, sampled-attention framing of Visual Occlusion Attentional Demand.

Key Points

Part 1 found a visual cost mediated by scanning, supporting multiple resources over preemption. Moving information from the auditory to the visual channel in an already visually loaded cockpit was predicted to cut both ways: multiple-resource theory predicts a cost to the high-priority visual tasks when a side task is also visual, whereas preemption theory predicts a benefit of visual delivery because a discrete auditory onset would otherwise capture attention away from flying. The data favoured multiple resources — vertical tracking and traffic detection were better under auditory side-task delivery, and the cost of visual delivery was traced to the reduction in outside-world scanning imposed by looking down at the CDTI, including a period of "horizon deprivation" that degraded flight-path control (PDF pp. 13–14, orig. pp. 371–372). Redundant audio-plus-visual delivery produced no "best of both worlds": it never outperformed the better single modality and sometimes gave the worst performance, behaving as an "attention sink" unless pilots are trained to manage it (PDF p. 14, orig. p. 372).

The expected-value model treats scanning as driven by expectancy and value. Wickens et al. frame visual sampling with four factors — salience (S), effort (E), expectancy (E), and value (V) — but argue that an optimal scanner should ignore salience and not let effort inhibit useful long scans, so the optimal model rests only on the expectancy and value terms; the full four-term form is the SEEV model (Wickens et al., 2001) (PDF pp. 14–15, orig. pp. 372–373). Expectancy that a location holds new information is operationalised as the bandwidth (rate of change) of each area of interest, after Senders (1964); value is the importance of the task the area serves, set by the aviate-navigate-communicate priority hierarchy (Schutte & Trujillo, 1996). The predicted dwell on an area of interest sums, across every task it serves, the product of its bandwidth, its relevance to that task, and that task's value, and is then expressed as a proportion of the total across areas (PDF pp. 15–16, orig. pp. 373–374). Coefficients are assigned by simple ordinal rank-ordering (e.g. aviate = 3, navigate = 2, communicate = 1), which lets modellers reach consensus on the inputs without precise measurement (PDF p. 16, orig. p. 374).

The model fit and cross-validated well, with bandwidth the dominant term. Fit against percentage-dwell-time data in an initial free-flight experiment, the three-parameter model gave a linear correlation of r = .885 (78% of variance), much of it carried by three clusters of dwell on the instrument panel, outside world, and CDTI (PDF p. 17, orig. p. 375). Cross-validated on the independent traffic data of Part 1 it reached R² = 90% (r = .93), and on a third communications experiment (Helleberg & Wickens, 2003) it accounted for 95% of the variance (PDF pp. 17–18, orig. pp. 375–376). Stripping parameters showed that bandwidth (expectancy) dominated the prediction, with weaker independent contributions from relevance and task priority. At the individual level the model fit each pilot's scanning with correlations from .42 to .95 (mean .85), and the goodness of fit rose with the pilot's flight experience (r = .55), echoing the long-standing finding that expert scanning is closer to optimal (PDF p. 18, orig. p. 376).

Conclusion

Wickens et al. (2003) conclude that introducing visual cockpit displays in place of ATC voice carries a real, if modest, cost to the highest-priority tasks of aviating and navigating — a caution for single-pilot operations — because head-down scanning to the new displays steals time from the instrument panel and outside world. The deeper contribution is methodological: an optimal expected-value model resting on expectancy (bandwidth) and value (task priority) predicts where pilots look, accounting for 78–95% of the variance in percentage dwell time across three independent data sets, with expectancy the strongest driver and fit improving with expertise. For this knowledge base the work supplies a validated, gaze-based account of selective visual attention — a computational "gold standard" for optimal scanning against which trainees and interface designs can be measured, and the direct aviation predecessor to the cognitive driver model in Attention Allocation Driver Model.

References

Helleberg, J.R. and Wickens, C.D. (2003) 'Effects of data link modality and display redundancy on pilot performance: An attentional perspective', The International Journal of Aviation Psychology, 13(3), pp. 189–210. To be validated.

Schutte, P.C. and Trujillo, A.C. (1996) 'Flight crew task management in non-normal situations', Proceedings of the Human Factors and Ergonomics Society Annual Meeting, 40(4), pp. 244–248. To be validated.

Senders, J.W. (1964) 'The human operator as a monitor and controller of multidegree of freedom systems', IEEE Transactions on Human Factors in Electronics, HFE-5(1), pp. 2–5. To be validated.

Wickens, C.D., Goh, J., Helleberg, J., Horrey, W.J. and Talleur, D.A. (2003) 'Attentional Models of Multitask Pilot Performance Using Advanced Display Technology', Human Factors, 45(3), pp. 360–380. doi: 10.1518/hfes.45.3.360.27250. wickens2003attentional

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. Savoy, IL: University of Illinois, Aviation Research Lab. To be validated.

Open Questions

  • The model is fit by assigning ordinal coefficients (bandwidth, relevance, value) by judgement. Whether empirically measured event rates from eye-tracking data could replace the hand-set bandwidth term — the same question raised for the AIE model in Attention Allocation Driver Model — is open.
  • Parameter-removal showed bandwidth (expectancy) dominating, with weak independent effects of value/priority. How far the value term matters when task priorities are less culturally fixed than aviate-navigate-communicate is untested here.
  • The optimal model deliberately drops salience and effort. In VR/HMD viewing, where saccade/head-rotation effort is larger and salient peripheral events are common, whether the two dropped terms re-enter as significant predictors is an open question for this KB's contexts.