Active Sensing Shapes Real-World Decision-Making through Dynamic Evidence Accumulation

TL;DR

A computational scheme is proposed to formalize real-world evidence affordance and capture active sensing through eye movements and provides a comprehensive account of how active sensing underlies real-world decision-making, unveiling multifactorial, integrated characteristics in real-world decision-making.

Abstract

Human decision-making heavily relies on active sensing, a well-documented cognitive behaviour for evidence gathering to accommodate ever-changing environments. However, its operational mechanism in the real world remains non-trivial. Currently, an in-laboratory paradigm, called evidence accumulation modelling (EAM), points out that human decision-making involves transforming external evidence into internal mental beliefs. However, the gap in evidence affordance between real-world contexts and laboratory settings hinders the effective application of EAM. Here we generalize EAM to the real world and conduct analysis in real-world driving scenarios. A cognitive scheme is proposed to formalize real-world evidence affordance and capture active sensing through eye movements. Empirically, our scheme can plausibly portray the accumulation of drivers' mental beliefs, explaining how active sensing transforms evidence into mental beliefs from the perspective of information utility. Also, our results demonstrate a negative correlation between evidence affordance and attention recruited by individuals, revealing how human drivers adapt their evidence-collection patterns across various contexts. Moreover, we reveal the positive influence of evidence affordance and attention distribution on decision-making propensity. In a nutshell, our computational scheme generalizes EAM to real-world contexts and provides a comprehensive account of how active sensing underlies real-world decision-making, unveiling multifactorial, integrated characteristics in real-world decision-making.

Figures (6)

• Figure 1: Overall formalization. (a) Illustration of how active sensing comes into play in real-world driving. (b) Two types of real-world driving scenarios: lane-changing and car-following. (c) An overview of our findings.
• Figure 2: The mental beliefs of drivers in different lane-changing scenarios portrayed by DEAM. (a) First-person view scenarios and distribution of momentary evidence for a lane-changing decision portrayed by DEAM. We plot the distribution of momentary evidence portrayed by DEAM, received by drivers at each time step related to lane-changing decision-making. The purple triangular scatters represent the momentary evidence the driver receives from FV. The yellow circular scatters represent the momentary evidence from the RV. For this descriptive figure, we use the following parameters: $\sigma^2_z = 1, \theta = 0.3, \Delta t = 0.01$. To better illustrate the distribution of the driver's attention, we provide first-person view images of scenarios that human drivers receive in the upper panel and the heatmap representing the spatial distribution of the driver's visual attention during that period. (b) Dynamical computation of RDV signals over time portrayed by DEAM. The blue RDV curves and purple lines represent the driver's attention being allocated to the FV, while the yellow and green represent the attention allocated to the RV. The orange dashed lines in the figure represent the decision bounds of the DEAM, while the red circles denote the points at which the decision-making process is concluded. When the accumulated mental belief reaches the upper bound, it is interpreted as the driver having made the decision for lane-changing. Conversely, reaching the lower bound indicates lane-keeping decisions. For this descriptive figure, we use the following parameters: $d = 0.002$, $\sigma^2_\text{model} = 0.02^2$, $\theta = 0.3$, $\Delta t=0.001$. (c)First-person view scenarios and distribution of momentary evidence for a lane-keeping decision portrayed by DEAM. The blue triangular scatters represent the momentary evidence the driver receives from FV. The green circular scatters represent the momentary evidence from RV. We use the following parameters for this descriptive figure: $\sigma^2_z = 1, \theta = 0.3, \Delta t = 0.01$. The bottom panel shows the first-person view scenario of the decision-making process in which the driver's final decision is lane-keeping.
• Figure 3: The mental beliefs of drivers in car-following scenarios portrayed by DEAM. (a) First-person view scenarios and distribution of momentary evidence for a decelerating decision portrayed by DEAM. We present the distribution of momentary evidence received by drivers at each time step for deceleration decision-making, depicted by DEAM. Grey triangular scatters represent momentary evidence from surroundings, while yellow circular markers represent momentary evidence from FV. The parameters used for this descriptive figure are $\sigma^2_z = 1, \theta = 0.4, \Delta t = 0.01$. The upper panel illustrates a first-person view scenario of the decision-making process where the final decision is decelerating. (b) Dynamic computation of RDV signals over time modelled by DEAM. The grey RDV curves depict periods when the driver's attention is distributed among surrounding items, while the yellow lines represent periods when attention is allocated to FV. The upper orange dashed line serves as the decision bound for decelerating, with a red circle indicating the point where the decision-making process ends. When the cumulative mental belief reaches the upper bound, it indicates the driver opted for decelerating. Reaching the lower grey bound signifies a decision to maintain the current speed. For this descriptive illustration, the parameters used are $d = 0.008$, $\sigma^2_\text{model} = 0.02^2$, $\theta = 0.4$, and $\Delta t=0.001$.
• Figure 4: Replicating human driver behaviour using DEAM and aDDM in lane-changing and car-following scenarios. (a) The probability of lane-changing monotonically increases as a function of evidence bias. The bars represent the empirical data from human driver behaviour ($t(7)=21.8, p<0.001$); the thick pink dashed line represents the fitting results from DEAM ($t(7)=107.4, p<0.001$); the yellow dashed line represents the fitting results from aDDM ($t(7)=72.6, p<0.001$). Additionally, the thin pink dashed lines represent the DEAM fitting results for five other sets of different parameter values. (b) Decelerating probability increases consistently with evidence bias (human data: $t(5) = 5.02, p < 0.001$; DEAM: $t(5)=39.32, p<0.001$; aDDM: $t(5) = 33.32, p < 0.001$). (c) RT decreases monotonically as a function of evidence clarity (human data: $t(7)= -3.53, p<0.01$; DEAM: $t(7)=-8.56, p<0.001$; aDDM: $t(7)= -9.44, p<0.001$). Both human drivers and computational models show shorter RT when more clear evidence for decision-making is provided. (d) Comparative analysis of RT and evidence clarity in lane-changing (top) and lane-keeping (bottom) decisions for data, DEAM, and aDDM. (e) RT in car-following decision-making decreases as evidence clarity increases (human data: $t(5) = 3.03, p < 0.05$; DEAM: $t(5) = -11.82, p < 0.001$; aDDM: $t(5) = -13.20, p < 0.001$). (f) Comparative analysis of RT and evidence clarity in decelerating(top) and keep-driving (bottom) decisions for data, DEAM, and aDDM. (g) The number of attentional switches in lane-changing decision-making decreases with evidence clarity. For human drivers, the number of attentional switches correlates with evidence clarity ($p<0.001$) and shows a monotonically decreasing trend. For models, decrease in attentional switching number as a function of evidence clarity (DEAM: $t(7)=-2.4, p<0.05$; aDDM: $t(7)=3.46, p<0.05$). When the evidence for decision-making is clearer, the results show a decrease in the number of attentional switches. (h) Comparative analysis of attentional switching number and evidence clarity in lane-changing (top) and lane-keeping (bottom) choices for data, DEAM, and aDDM. (i) The number of attentional switching in car-following decreases with evidence clarity. For human drivers, the number of attentional switches correlates with evidence clarity ($p < 0.05$). DEAM demonstrates fewer attentional switches with clearer evidence (DEAM: $t(5) = -4.19, p < 0.01$). (j) Comparative analysis of attentional switching numbers and evidence clarity in decelerating(top) and keep-driving (bottom) decisions for data, DEAM, and aDDM.
• Figure 5: Attentional switching probability of human drivers and DEAM across different evidence clarity levels over time. The x-axis in the 3D plot represents elapsed time, the y-axis depicts evidence clarity, and the z-axis represents the smoothed switching probability. The surface is projected on the z=-0.02 plane. Insets in the top right corner of the 3d plot separately display the overall attentional switching probability (left insets), and attentional switching probability at the highest and lowest evidence clarity levels (right insets). (a) Attentional switching probability results of human drivers in lane-changing scenarios. (b) Attentional switching probability results of DEAM in lane-changing scenarios. (c) Attentional switching probability results of human drivers in car-following scenarios. (d) Attentional switching probability results of DEAM in car-following scenarios.