NOVA: Symbolic Regression Discovery of Interpretable Car-Following and Lane-Change Models with Driver Heterogeneity

NOVA, a symbolic regression framework, discovers interpretable models of human driving behavior from 4.7 million real-world observations, achieving superior performance on car-following and lane-change prediction tasks. The research demonstrates that complex driving dynamics can be captured through compact algebraic structures that generalize across different freeway locations and driver populations.

NOVA represents a significant advancement in interpretable machine learning applied to autonomous vehicle development. Rather than relying on black-box neural networks, the framework uses symbolic regression to identify mathematically compact models that explain human driving behavior. Testing on nearly 5 million real driving observations from California highways, NOVA outperforms previous symbolic regression approaches by 9.8% in RMSE while maintaining interpretability—a critical advantage for safety-critical autonomous systems that regulators and insurers increasingly demand.

The research addresses a fundamental challenge in autonomous vehicle development: understanding and replicating human driving patterns with models transparent enough for validation and debugging. Previous approaches either sacrifice accuracy for interpretability or vice versa. NOVA's discovery of a robust two-term acceleration model suggests human car-following follows simpler rules than many neural network-based systems assume, with findings grounded in established psychophysical collision-avoidance theory.

The zero-shot transfer capability across freeway locations holds particular commercial value, suggesting these models could reduce the data collection and recalibration burden for deploying autonomous systems in new geographic regions. The 67.4% balanced accuracy on lane-change prediction, surpassing baselines by nearly 30 percentage points, demonstrates applicability beyond basic car-following to complex multi-agent scenarios.

For the autonomous vehicle industry, this work strengthens the case for explainable AI approaches and could inform regulatory frameworks requiring interpretable decision-making. Development teams may adopt similar symbolic regression methods to accelerate model validation and reduce liability exposure. The research also hints that human driving may be more systematically modeled than previously believed, potentially enabling more efficient training data collection.

• →NOVA discovers compact, interpretable driving models outperforming prior symbolic regression baselines by 0.135 m/s² RMSE on car-following prediction.
• →Models generalize zero-shot across freeway locations with minimal performance degradation, reducing geographic recalibration needs.
• →Lane-change prediction achieves 67.4% balanced accuracy on unseen drivers, 29.8 percentage points above existing baselines.
• →Discovered structures align with established psychophysical collision-avoidance theory, linking data-driven discovery to behavioral science.
• →Interpretable algebraic models offer regulatory and safety advantages over black-box neural networks for autonomous vehicle deployment.