Is trust a belief,observation,or state: Results from an active inference analysis of driver-automation transitions of control

Abstract

Transitions of control are an important safety concern for human-automation teams and automated vehicle safety. While trust and situation awareness have been observed to influence transitions of control in automated vehicles, there are few objective measurements, making these concepts difficult to operationalize in increasingly automated decision systems. In this study, we take a step towards quantifying trust by mapping latent driver beliefs extracted from an active inference-factor analysis model of driver behavior and cognitive dynamics to subjective responses to trust questionnaires. Our results show that subjective trust is primarily correlated with model parameters affecting perceptual evidence accumulation rate, and the same parameters are significantly correlated with driver age.

Keywords

active inference driver modeling trust situation awareness

Driver process models are an effective method for designing more effective vehicle automation (Bärgman et al., 2017; Roesener et al., 2017). Despite their broad applicability, current driver process models are limited by their ability to model internal driver decision making processes, i.e., cognitive dynamics. Active inference theory provides a promising platform for modeling these driver cognitive dynamics (Engström et al., 2022; Wei, Garcia, et al., 2022; Wei, McDonald, et al., 2022), however the relationship between modeled and measured cognitive dynamics is not well established. The goal of this work is to establish this relationship by fitting an active inference model and analyzing correlations between model parameters and subjective trust, situation awareness, fatigue, and demographics.

We accomplished this goal by fitting active inference models to a driving simulation dataset including 38 participants (23 Males and 15 Females) aged 25-42 (M = 28.0 (3.4)). The driving simulation consisted of a construction zone scenario where the participant’s vehicle had the automation engaged and was following a lead vehicle and approaching a straight roadway partially obstructed by traffic cones. Driver behaviors were modeled with an active inference driver model in the form of a partially observable Markov decision process (POMDP) with 2 observations (the distance to the cones and time since last automation disengagement (TSLC)), 2 actions (engage, disengage the automation), and 2 states. A factor analysis was performed on the fitted models to identify meaningful variations in the fitted parameters. A subsequent correlation analysis was used to understand relationships between the factors and the subjective factors.

The results illustrated a close mapping between predicted and actual driver takeover times and a K-S test comparing reaction time distributions failed to reject the null hypothesis that the distributions are identical. The factor analysis identified 3 factors that sufficiently explained variance in the parameters representing driver preferences and uncertainty about the environment (Factor 1), decreased expectation of obstacle-free distance, increased expectation of obstacle-free TSLC, and decreased expectation of obstacle-present TSLC (Factor 2), higher state-transition probabilities, especially regarding the obstacle-present state when disengaged (Factor 3). Factor 1 is correlated with Situation Awareness (SA; 0.1), and changes in SA before and after the event (Delta SA; -0.18). Factor 2 is correlated with Sex, Age, Driving years (0.17), Dynamic Trust (DT; 0.11), SA (-0.18), and Delta SA (-0.13). Factor 3 is correlated with DT, Sex, Age, and Driving years. These results suggest that driver SA is mainly correlated with preference (factor 1) and observations parameters (factor 2) which affect driver action selection, while driver dynamics trust is mainly correlated with observation (factor 2) and transition (factor 3) parameters which affect driver beliefs. Our results show that subjective trust is mostly associated with driver mental model affecting perceptual evidence accumulation rate (i.e., driver beliefs). We encourage future work to replicate the current results on a larger sample size and further investigate the connections between trust and beliefs.

Footnotes

Acknowledgements

This work was partially supported by a grant from the National Science Foundation (Award#: 2035367).

ORCID iD

Anthony D. McDonald

References

Bärgman

Boda

C.-N.

Dozza

(2017). Counterfactual simulations applied to SHRP2 crashes: The effect of driver behavior models on safety benefit estimations of intelligent safety systems. Accident Analysis&Prevention, 102,165–180. https://doi.org/10.1016/j.aap.2017.03.003

Engström

Liu

S.-Y.

Dinparastdjadid

Simoiu

(2022). Modeling road user response timing in naturalistic settings: A surprise-based framework (arXiv:2208.08651).arXiv.https://doi.org/10.48550/arXiv.2208.08651

Roesener

Hiller

Weber

Eckstein

(2017). How safe is automated driving? Human driver models for safety performance assessment. 2017 IEEE 20th International Conference on Intelligent Transportation Systems (ITSC), 1–7. https://doi.org/10.1109/ITSC.2017.8317706

Wei

Garcia

McDonald

Markkula

Engström

Supeene

O’Kelly

(2022). World Model Learning From Demonstrations With Active Inference: Application to Driving Behavior. Forthcoming, 3rd International Workshop on Active Inference, Grenoble, France.

Wei

McDonald

A. D.

Garcia

Alambeigi

(2022). Modeling driver responses to automation failures with active inference. IEEE Transactions on Intelligent Transportation Systems, 23(10), 18064–18075.