Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Learning An Active Inference Model of Driver Perception and Control: Application to Vehicle Car-Following

Ran Wei, Anthony D. McDonald, Alfredo Garcia, Gustav Markkula, Johan Engstrom, Matthew O'Kelly

TL;DR

This work proposes a Bayesian MAP framework to learn a perception-and-control model for human-like driving within an active-inference-inspired POMDP. By parameterizing perception and preferences separately and solving a bi-level optimization, the authors train an interpretable model (AIDA) that competes with BC and IDM on car-following data from the INTERACTION dataset. Across offline and online evaluations, AIDA shows competitive accuracy and offers clearer interpretability through its belief-state and input-output mappings, though it faces challenges in extreme scenarios due to limited data coverage. Overall, the study demonstrates that data-driven active inference models can provide a transparent, cognitively grounded alternative to black-box driving models with meaningful implications for human-centric autonomy and safety.

Abstract

In this paper we introduce a general estimation methodology for learning a model of human perception and control in a sensorimotor control task based upon a finite set of demonstrations. The model's structure consists of i the agent's internal representation of how the environment and associated observations evolve as a result of control actions and ii the agent's preferences over observable outcomes. We consider a model's structure specification consistent with active inference, a theory of human perception and behavior from cognitive science. According to active inference, the agent acts upon the world so as to minimize surprise defined as a measure of the extent to which an agent's current sensory observations differ from its preferred sensory observations. We propose a bi-level optimization approach to estimation which relies on a structural assumption on prior distributions that parameterize the statistical accuracy of the human agent's model of the environment. To illustrate the proposed methodology, we present the estimation of a model for car-following behavior based upon a naturalistic dataset. Overall, the results indicate that learning active inference models of human perception and control from data is a promising alternative to black-box models of driving.

Learning An Active Inference Model of Driver Perception and Control: Application to Vehicle Car-Following

TL;DR

This work proposes a Bayesian MAP framework to learn a perception-and-control model for human-like driving within an active-inference-inspired POMDP. By parameterizing perception and preferences separately and solving a bi-level optimization, the authors train an interpretable model (AIDA) that competes with BC and IDM on car-following data from the INTERACTION dataset. Across offline and online evaluations, AIDA shows competitive accuracy and offers clearer interpretability through its belief-state and input-output mappings, though it faces challenges in extreme scenarios due to limited data coverage. Overall, the study demonstrates that data-driven active inference models can provide a transparent, cognitively grounded alternative to black-box driving models with meaningful implications for human-centric autonomy and safety.

Abstract

In this paper we introduce a general estimation methodology for learning a model of human perception and control in a sensorimotor control task based upon a finite set of demonstrations. The model's structure consists of i the agent's internal representation of how the environment and associated observations evolve as a result of control actions and ii the agent's preferences over observable outcomes. We consider a model's structure specification consistent with active inference, a theory of human perception and behavior from cognitive science. According to active inference, the agent acts upon the world so as to minimize surprise defined as a measure of the extent to which an agent's current sensory observations differ from its preferred sensory observations. We propose a bi-level optimization approach to estimation which relies on a structural assumption on prior distributions that parameterize the statistical accuracy of the human agent's model of the environment. To illustrate the proposed methodology, we present the estimation of a model for car-following behavior based upon a naturalistic dataset. Overall, the results indicate that learning active inference models of human perception and control from data is a promising alternative to black-box models of driving.
Paper Structure (31 sections, 2 theorems, 41 equations, 12 figures, 6 tables, 1 algorithm)

This paper contains 31 sections, 2 theorems, 41 equations, 12 figures, 6 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. A POMDP Model of Perception and Control
  3. Estimation Methodology
  4. An Active Inference Specification
  5. Application: Learning a Model of Perception and Control in Car Following Behavior
  6. Models and parameterization
  7. Dataset
  8. Feature Computation
  9. Model Evaluation and Comparison
  10. Offline Evaluation
  11. Online Evaluation
  12. Statistical Evaluation
  13. Results and Discussion
  14. Offline Performance Comparison
  15. Offline Performance Comparison
  16. ...and 16 more sections

Key Result

Proposition II.1

Let $V_{t}(b)$ be recursively defined as follows: where $b'(s) =\mathbb{P}(s_{t+1}=s|h_t\cup(a,o'))$, i.e. the resulting Bayes update after action $a$ is implemented and observation $o'$ are recorded. Then, the Bayes updated belief $b_t=\mathbb{P}(\cdot|h_t)$ is a sufficient statistic for solving model, i.e. $U_{t}(h_t)=V_{t}(b_t)$ for all $h_t$.

Figures (12)

  • Figure 1: Graphical Model of Perception and Control.
  • Figure 2: Top-down view of the roadway explored in the analysis. The west-bound lanes (blue) have denser traffic and more stop-and-go behavior whereas the east-bound lanes (orange) have sparser traffic and higher speed. We trained the models to emulate the behavior of the blue cars and evaluated the models’ ability to predict the behavior of the blue and orange cars. Grey cars in the merging lanes were excluded.
  • Figure 3: Offline evaluation MAE-IQM. Each point corresponds to a random seed used to initialize model training and its color corresponds to the testing condition of either dense-lane or sparse-lane.
  • Figure 4: Example offline predictions in the dense-lane (top) and sparse-lane (bottom) settings. Each line except for the blue line represents the mean prediction of the corresponding model. Shading represents 1 standard deviation of prediction interval. The prediction intervals for BC and AIDA are computed by drawing 30 samples from the models' predictive distributions. IDM has no prediction interval because it's deterministic.
  • Figure 5: Online evaluation ADE-IQM. Each point corresponds to a random seed used to initialize model training and its color corresponds to the testing condition of either dense-lane or sparse-lane.
  • ...and 7 more figures

Theorems & Definitions (5)

  • Proposition II.1
  • proof
  • Theorem II.2
  • proof
  • proof