An Active Inference Model of Car Following: Advantages and Applications

Wei, Ran; McDonald, Anthony D; Garcia, Alfredo; Markkula, Gustav; Engström, Johan; O’Kelly, Matthew · 2023 · ROSA P / arXiv [Cornell University]

archive: archived pipeline: cataloged verified

Get this paper ↗ (full text — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the need for driver behavior models that are simultaneously flexible, generalizable, and interpretable for the development of automated vehicles. Existing rule-based models, such as the Intelligent Driver Model (IDM), lack the flexibility to mimic nuanced social behaviors, while data-driven models like Behavior Cloning (BC) offer flexibility but suffer from poor interpretability and reliance on large datasets. To bridge this gap, the authors propose the Active Inference Driving Agent (AIDA), a model grounded in active inference theory. This framework posits that agents maintain internal probabilistic generative models of their environment and select actions to minimize expected free energy, thereby balancing goal-seeking and information-seeking behaviors. The study evaluates AIDA against the rule-based IDM and two neural network BC models (Multi-Layer Perceptron and Recurrent Neural Network). The models were trained and tested using a subset of the INTERACTION dataset, which contains real-world driving trajectories from a highway segment in China. The dataset was filtered to include 1,254 car-following trajectories, with features including distance headway, relative velocity, and a perceptual control analog of inverse time-to-collision. Model performance was assessed through offline evaluations, measuring the Mean Absolute Error (MAE) of predicted acceleration, and online evaluations using a simulator to measure Average Deviation Error (ADE) and Lead Vehicle Collision Rate (LVCR). Testing was conducted on both "same-lane" data (unseen vehicles in familiar conditions) and "new-lane" data (unseen vehicles in novel traffic conditions with higher speeds and lower density). Results indicate that AIDA significantly outperformed the IDM in both offline prediction accuracy and online trajectory similarity across all testing conditions. In offline evaluations, AIDA achieved the lowest MAE in same-lane tests and performed comparably to BC models in new-lane tests. In online simulations, AIDA demonstrated significantly better trajectory matching than IDM and BC-RNN, though BC-MLP achieved slightly lower deviation errors. Collision rates were low across all models, with AIDA showing variability dependent on random seeds but maintaining minimum collision rates comparable to baselines. Crucially, interpretability analyses revealed that AIDA’s learned distributions aligned with established driver behavior theory. Visualizations of the model’s internal beliefs allowed researchers to trace decision-making processes and identify errors attributable to limited training data, a capability absent in black-box BC models. The significance of this work lies in demonstrating that active inference can provide a viable alternative to opaque data-driven models by combining their behavioral flexibility with the interpretability of rule-based systems. The findings suggest that AIDA can effectively model complex car-following behaviors while allowing for diagnostic inspection of internal states. The authors conclude that further research should focus on modeling diverse driving styles and training with more varied datasets to enhance generalizability, highlighting the potential of psychologically grounded probabilistic frameworks in autonomous vehicle development.

Key finding

The Active Inference Driving Agent significantly outperformed the rule-based Intelligent Driver Model in predicting driving controls and achieved accuracy comparable to neural network Behavior Cloning models while maintaining interpretability grounded in driver behavior theory.

Methodology

modeling

Sample size: 1254

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed. Discovered via bulk_ingest_rosap on 2026-05-23 (6 acquisition events logged).

Summary generated by qwen3.6-27b-prismaquant on 2026-06-10; verification: verified_with_issues.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• mental model of traffic
• situational awareness
• anticipation

Information type

What kind of knowledge this paper contributes, grouped by family — independent of topic (what it is about) and method (how it was studied).

• Methodological Resource: tool software
• Theoretical Contribution: computational model, theory or model