Fig 1.
Samples of hypothetical exit-choice scenarios.
Each scenario offers four alternative exits whose attributes compete with one another other. The decision maker is asked to make a trade-off between those factors in each scenario and choose the exit that they would choose in an evacuation scenario. (A) Sample of a hypothetical scenario in which two exits are invisible to the decision maker from their current position due to the presences of obstacles (a choice scenario with partially-ambiguous attributes for a subset of the choice set). The invisible exits are presented in a blurred way indicating that the level of congestion around those alternatives is not given to the decision maker. (B) Sample of a scenario in which all four exits are visible from the current position of the decision maker (a fully unambiguous choice scenario).
Fig 2.
Snapshots from the raw footage of the mock evacuation experiments.
A camera at 8m height above the floor records the experiments. The obstacle inside the room creates exit invisibility similar to the conditions in the actual building represented by the artificial model. Participants start evacuating at the entrance and are asked to exit the model building as quick as possible through one of the exits available to them. (A) A lightly congested treatment (75 evacuees). (B) A congested treatment (150 evacuees).
Fig 3.
Raw footage of the trials is analyzed using specialty software calibrated for the specific environmental conditions in the field.
Individuals are recognized and tracked by the software based on the color of the beanies they wore during the experiments. Trajectory, body movement and head orientation of each participant are inspected as external indicators to identify their likeliest decision moments. Decision moments are extracted as the moments after which participants showed most certainty and consistency in their movements. (A) A subject who chose exit 2, at their identified decision moment. (B) A subject who chose exit 3, at their identified decision moment.
Fig 4.
Extracting RC observations from the field-type evacuation experiments.
A subject is shown at their decision moment along with their full trajectory of movement. The choice set and attribute levels of all alternative exits are recorded at the moment of the decision which together with the chosen alternative (here, Exit 2) constitute a single choice observation. The attributes that are recorded include C1–C4 that signify congestion (CONG) around each exit (the number of evacuees congregating near each exit), F1–F4 that denote the size of directional flows (the number of evacuees) moving to each exit (FLTOEX), V1–V4 as binary 0–1 variables representing visibility status of each exit (which equals 1 if the exit is visible from the position that the decision is made), D1–D4 that represent the subject’s distance to each exit, calculated based on the coordinates of the subject at their decision moment and measured in meters.
Table 1.
Model estimation results on HC and RC datasets, and the combined (RC+HC) data.
Fig 5.
Comparing the model estimates obtained from the HC against RC data.
The error bars represent the 95% confidence interval of each estimate. The red square separates the HC-based estimates from those obtained from the RC observations. (A) Estimates obtained from a FP-MNL model specification. (B) Estimates obtained from a RP-MNL model specification (the mean of utility coefficients). Note that according to the scale difference of the models derived from the HC and RC segments, comparison between the corresponding “magnitude” of the individual estimates between the HC and RC contexts would not be meaningful and is not intended by these graphs. Only, the sign, significance and the estimate patterns of the parameters are meaningfully comparable here.
Fig 6.
The impact of the context dependency (HC versus RC) and preference heterogeneity on the model predictions.
Each point in the scatterplots (A)-(D) is associated with one alternative exit in one choice situation extracted from the field-type trials. The coordinates of each point are the probabilities predicted for that alternative (exit) by the model represented by the two axes. Figures (A) and (B) measure the effect of context dependency on predictions by making pairwise comparisons between the predictions of the counterpart models obtained from the real and hypothetical experiments. Figures (C) and (D) measure the effect of accommodating preference/perception heterogeneity on the predictions by making pairwise comparisons between the predictions obtained from models that can and cannot accommodate the heterogeneity effect (estimated on the same data). The average of absolute differences (Avg diff.) and the correlations (Corr) between each pair of data series are shown on each scatter plot. Figure (E)-(H) color-code the concentration of the points (the number of points per 1×1 square) in the scatterplots presented in Figures (A)-(D) respectively. Figures (I)-(L) show the distribution of the absolute differences between the predicted probabilities of each model pair compared in figures (A)-(D) respectively (i.e. the distribution of the absolute differences between the values of the x and y coordinates of the points shown in figures (A)-(D)).