Enhancing Automated Vehicle Safety Through Testing with Realistic Driver Models

Garcia, Alfredo; McDonald, Anthony D · 2023 · ROSA P / Safety through Disruption (Safe-D) University Transportation Center (UTC)

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 research addresses the challenge of developing driver behavior models that are simultaneously flexible, generalizable, and interpretable for the testing and verification of automated vehicles. Existing approaches are divided into rule-based models, such as the Intelligent Driver Model (IDM), which are interpretable but lack behavioral flexibility, and data-driven machine learning models, which offer flexibility but suffer from poor interpretability and high data requirements. To bridge this gap, the authors developed the Active Inference Driving Agent (AIDA), a novel car-following model based on the active inference framework. This approach leverages Bayesian principles of cognition, allowing the agent to maintain internal probabilistic models of the environment to minimize surprise and uncertainty, thereby combining the structural interpretability of rule-based systems with the adaptive flexibility of data-driven methods. The study implemented AIDA using a Partially Observable Markov Decision Process (POMDP). The model was specified to observe relative velocity, distance headway, and visual looming of the lead vehicle, with actions defined as continuous acceleration inputs. Model parameters were learned from the INTERACTION dataset, a real-world driving dataset containing trajectories from the USA, Germany, and China, using maximum likelihood estimation. AIDA was benchmarked against three comparison algorithms: the rule-based IDM and two data-driven behavior cloning models (BC-MLP and BC-RNN). Performance was evaluated through offline trajectory prediction accuracy and online control deviation, using metrics such as Mean Absolute Error (MAE) and Average Deviation Error (ADE). Results indicated that AIDA significantly outperformed the IDM in offline prediction accuracy and achieved comparable accuracy to neural network models in three out of four evaluations. In online control settings, AIDA demonstrated lower average deviation errors than both IDM and BC-RNN. Crucially, interpretability analyses revealed that AIDA’s learned distributions aligned with established driver behavior theory. Visualizations of the model’s belief and action distributions allowed researchers to directly comprehend decision-making processes and diagnose errors, a capability lacking in opaque neural network models. However, AIDA showed sensitivity to initial training conditions, resulting in higher collision rates in certain scenarios compared to baseline models. The significance of this work lies in demonstrating that active inference can provide a viable middle ground between rigid rule-based models and uninterpretable data-driven models. By grounding the model in psychological theory, AIDA offers superior transparency for model inspection and verification, which are critical for automated vehicle safety. The findings suggest that while AIDA matches the predictive performance of complex neural networks, its structural interpretability enables better diagnostic analysis. The authors recommend future work to expand observation modalities and incorporate domain knowledge to address current limitations regarding dataset coverage and model sensitivity to local optima.

Key finding

The Active Inference Driving Agent significantly outperformed rule-based models in predicting driving controls while maintaining comparable accuracy to data-driven neural networks and offering superior interpretability through visualizable belief distributions.

Methodology

modeling

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.

• situational awareness
• mental model of traffic
• automation surprise
• 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