🤖 machine learning

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

The paper introduces NOVA, an autonomous symbolic regression framework that discovers compact, interpretable car-following and lane-change models from massive trajectory data, significantly outperforming existing baselines in prediction accuracy and generalization while revealing robust nonlinear structures linked to established psychophysical theories of human driving behavior.

Original authors: Ishak Abassi, Nassim Ali Bouazzouni, Farah Ibelaiden, Nadir Farhi

Published 2026-06-10

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ishak Abassi, Nassim Ali Bouazzouni, Farah Ibelaiden, Nadir Farhi

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a robot how to drive a car. For decades, scientists have tried to do this by writing down strict rules based on their own theories, like "if the car ahead is close, brake hard." But human drivers are messy, emotional, and unpredictable. They don't always follow the rules we think they should.

This paper introduces NOVA, a new kind of "scientific detective" that doesn't guess the rules. Instead, it watches millions of real driving videos and tries to figure out the hidden math that describes how humans actually behave. Think of it as a chef who doesn't follow a recipe book but instead tastes thousands of dishes to discover the exact secret ingredient that makes them delicious.

Here is what NOVA discovered, explained simply:

1. The "Sponge" Effect of Speed Differences

The most important discovery is a specific mathematical shape called a hyperbolic tangent (or tanh).

The Analogy: Imagine a sponge. When you squeeze it gently, it squishes a little. When you squeeze it hard, it squishes a lot. But once it's fully compressed, no matter how hard you squeeze, it can't get any smaller.
The Discovery: NOVA found that drivers react to the speed difference between them and the car ahead exactly like this sponge. If the car ahead is slightly faster or slower, the driver adjusts gently. But if the speed difference gets huge, the driver's reaction hits a "ceiling." They don't slam the brakes infinitely harder; they hit a maximum limit. Previous models assumed drivers reacted in a straight line (more speed difference = proportionally more braking), but NOVA proved humans have a "sponge-like" limit to their reactions.

2. The "Optical Looming" Safety Signal

NOVA also found that drivers use a specific visual trick to avoid crashes, which scientists call "optical looming."

The Analogy: Think of a fly buzzing toward your face. As it gets closer, it doesn't just get bigger; it seems to expand faster and faster in your vision. That rapid expansion is the "looming" signal.
The Discovery: NOVA found that drivers use this "expansion rate" (how fast the car ahead looks like it's growing in their windshield) to decide when to brake or change lanes. This wasn't programmed into the computer; the computer found it on its own. It turns out this same visual signal is used for both braking (car-following) and changing lanes, acting as a universal "danger alarm" for the human brain.

3. The "Driver Personality" Dial

The researchers realized that not all drivers are the same. They created a "personality dial" to measure how aggressive a driver is.

The Analogy: Imagine a volume knob on a stereo. Some drivers have the volume turned up (aggressive), some are at a medium level (normal), and some are turned down low (conservative).
The Discovery: NOVA found that while the basic "sponge" rule applies to everyone, the volume of the reaction changes based on personality. Aggressive drivers turn the volume up, reacting more intensely to speed differences, while conservative drivers keep it low. The computer could predict this personality just by watching how they drove.

4. The "Lane Change" Puzzle

The team also asked NOVA to figure out why people change lanes.

The Analogy: Changing lanes is like deciding whether to jump from one moving train to another. You need to know if the gap is big enough and if the other train is coming too fast.
The Discovery: NOVA built a model that is much better at predicting lane changes than the old standard models (which were only right about 37% of the time). NOVA got it right about 67% of the time. It found that drivers are very sensitive to the inverse of the gap (meaning, the closer the car, the exponentially higher the risk).
The Limit: However, the paper notes a ceiling. Even NOVA struggles to predict why a driver chooses to stay in their lane versus changing, especially when the situation is ambiguous. It seems that looking at just one snapshot in time isn't enough to understand the full story of a lane change; you need to see the history of what happened a few seconds ago.

5. The "Universal Translator"

Finally, the team tested if these rules work in different places.

The Analogy: If you learn a language in New York, can you speak it in Los Angeles without relearning it?
The Discovery: Yes. The mathematical rules NOVA found on one highway (I-80) worked almost perfectly on a completely different highway (US-101) without any adjustments. This proves that the "sponge" reaction and the "looming" danger signal are fundamental parts of how humans drive, not just quirks of a specific road.

Summary

In short, NOVA is a tool that stripped away human bias and found that human driving is governed by simple, discoverable math:

We react to speed differences like a sponge (with a limit).
We watch the expansion of the car ahead to judge danger.
We have personality knobs that turn our reactions up or down.
These rules work everywhere, regardless of the specific highway.

The paper concludes that while we can't perfectly predict every single lane change yet, we have finally found the core "grammar" of human driving behavior.

Technical Summary: NOVA – Symbolic Regression Discovery of Interpretable Car-Following and Lane-Change Models

1. Problem Statement

The paper addresses the challenge of discovering interpretable, compact mathematical structures that describe human driving behavior from raw trajectory data. While classical car-following models (e.g., IDM, Krauss) impose fixed functional forms determined by researchers rather than data, and deep learning approaches offer high predictive accuracy but lack interpretability (acting as "black boxes"), there remains a gap for a method that autonomously recovers symbolic equations without structural priors.

The specific scientific problem is to determine whether human car-following and lane-changing behaviors contain low-complexity symbolic structure that can be recovered from noisy, high-dimensional trajectory data (specifically the NGSIM dataset) without assuming specific functional forms (e.g., linear, polynomial, or specific saturation functions) in advance.

2. Methodology

2.1. The NOVA Framework

NOVA is an autonomous symbolic regression framework implemented in Rust for high-performance, deterministic execution. Unlike stochastic genetic programming approaches, NOVA employs a deterministic exhaustive combinatorial search over a bounded library of algebraic structures.

Search Space: The engine evaluates over 10,000 candidate algebraic structures derived from a foundational set of primitive kinematic atoms (velocity $v$ , leader velocity $v_l$ , acceleration $a_l$ , relative velocity $\Delta v$ , gap) and their temporally lagged counterparts.
Constraints: The search is restricted to polynomial degrees $\le 2$ to prevent overfitting to sensor noise and to align with physical kinematic principles. Crucially, no legacy model structures (e.g., IDM's free-road term) are injected as priors. Higher-order cognitive metrics like Time Headway (THW) or Time-to-Collision (TTC) are not pre-included; the engine must synthesize them from primitives (e.g., discovering $gap/v$).
Optimization: Coefficients are computed via closed-form Ordinary Least Squares (OLS) using pre-factored Singular Value Decomposition (SVD) for stability.

2.2. Data and Preprocessing

The study utilizes 4,765,788 active driving observations from the NGSIM I-80 and US-101 datasets.

Filtering: Strict quality filters remove coasting/idling ( $|a| < 0.2$ m/s²) and out-of-bounds kinematics.
Smoothing: To mitigate 10 Hz tracking jitter (which causes non-physical acceleration spikes upon double differentiation), two complementary preprocessing pipelines are used:
1. Pipeline R: Rolling mean (15-frame centered) with a 0.8 s forward-shifted target.
2. Pipeline S: Savitzky–Golay filter (order 3, window 15) with a 1.0 s mean target.
Evaluation Protocol: A strict 80/20 vehicle-ID holdout is enforced to ensure generalization to unseen drivers rather than memorization of specific trajectories.

2.3. Discovery Process

The discovery follows a multi-stage "residual cascade":

Phase 1 (Global Law): Discovery of the primary structure from raw data.
Phase 2 (Heterogeneity): Modeling driver personality (aggressiveness) and speed regimes (stop-and-go, transition, free-flow).
Phase 3 (Residuals): A second search on the residuals of Phase 2 to identify remaining structural signals (e.g., collision avoidance).

3. Key Contributions and Findings

3.1. Car-Following: The Invariant Backbone

The primary discovery is that the hyperbolic tangent of relative velocity, $\tanh(\Delta v)$ , is the invariant, dominant term in human car-following.

The Rank-4 Law: The top-ranked model is a two-term equation: $\hat{a} = -0.468 + 1.266 \tanh(\Delta v) + 0.194 \text{THW}$ .
Interpretation: The $\tanh(\Delta v)$ term captures a bounded, saturating response to relative speed. For small relative speeds, the response is linear; for large speeds, acceleration saturates at $\approx \pm 1.266$ m/s², consistent with psychophysical Weber–Fechner principles. This structure was not present in prior models (e.g., OVM applies saturation to spacing, not closure speed).
Robustness: This backbone was recovered independently across eight different experimental configurations (varying pipelines, horizons, driver populations, and feature timings), confirming it is not a preprocessing artifact.
Driver Heterogeneity: A continuous aggressiveness index ( $\kappa$ ) derived from a Gaussian Mixture Model (GMM) modulates the $\tanh(\Delta v)$ coefficient. Aggressive drivers show a stronger response to relative velocity, particularly in the "transition" speed regime.
Residual Discovery (TTC): A second-stage search on residuals identified $\tanh(\text{TTC})$ as the dominant remaining term. This connects the data-driven discovery to Lee's $\tau$ -theory of optical looming, providing a bounded formulation of collision risk without explicitly supplying optical-flow models as priors.

3.2. Lane-Change Modeling

NOVA was extended to a Multinomial Logit (MNL) framework to discover lane-change utility functions.

Performance: The discovered 16-feature model achieved 67.4% balanced accuracy on 502 unseen drivers under strict vehicle-ID holdout. This surpasses the standard MOBIL baseline (37.6%) by +29.8 percentage points and a hand-engineered MNL baseline by ~17 pp.
Structure: The utility functions are directionally decoupled. Key features include inverse gap ($1/gap$) and inverse time-to-collision ( $1/\text{TTC}$ ), confirming that optical looming is a shared safety signal across both longitudinal (braking) and lateral (lane-change) decisions.
Limits: Feature augmentation (adding more complex gap-acceptance physics) and residual boosting failed to improve out-of-driver accuracy. The paper identifies a structural ceiling for single-snapshot kinematic features in distinguishing "Stay" vs. "Change" decisions, suggesting future work requires temporal history features.

3.3. Cross-Location Transfer

The discovered symbolic operators demonstrated zero-shot transferability between the I-80 and US-101 sites.

Transferring the model trained on one site to the other resulted in only a 2.5–2.8 percentage point loss in $R^2$ , recovering 97–99% of the local oracle performance. This indicates the discovered structures are properties of human highway driving rather than dataset-specific fits.

4. Results and Benchmarks

Car-Following Accuracy: Under Pipeline S (Savitzky–Golay), NOVA achieved RMSE = 1.376 m/s² ( $R^2 = 15.57\%$ ), outperforming the best recalibrated symbolic-regression baseline (SR-LLM) by 0.135 m/s² in RMSE under an identical protocol.
Comparison to Classical Models: Classical models like IDM and Krauss performed poorly on instantaneous empirical snapshots (IDM $R^2 = -68\%$ on Pipeline R) because their fixed functional forms do not match the instantaneous acceleration dynamics of the NGSIM data.
Simulation Stability:
- Open-Loop: With a safety override (TTC < 1.5s), the 3-pass law achieved a 2.5% crash rate over 30-second horizons, outperforming calibrated IDM in long-horizon speed error.
- String Stability: A 13-vehicle platoon simulation yielded a stability ratio $\Lambda = 1.064$ and a backward wave speed of 35.4 m/s, matching empirical NGSIM observations of stop-and-go waves. This stability is an emergent property, not a design constraint.

5. Significance and Claims

The paper claims that NOVA successfully demonstrates that autonomous symbolic regression can recover interpretable structural regularities from noisy human behavioral data.

Scientific Discovery: The framework autonomously identified the $\tanh(\Delta v)$ nonlinearity and the $\tanh(\text{TTC})$ looming signal without these functional forms being supplied as priors. This suggests that human driving behavior contains low-complexity symbolic structure that was previously obscured by the fixed assumptions of classical models.
Convergence of Theory and Data: The data-driven discovery of optical looming ( $\text{TTC}$ ) in both car-following and lane-change contexts provides large-scale empirical evidence supporting Lee's $\tau$ -theory, unifying longitudinal and lateral safety signals.
Methodological Rigor: By enforcing strict vehicle-ID holdouts, multiple preprocessing pipelines, and zero-shot transfer tests, the paper argues that the discovered structures are robust behavioral invariants rather than statistical artifacts or overfitting to specific datasets.
Limitations Acknowledged: The authors modestly note that the model is specific to highway gap-following (failing on urban intersection data), that lane-change discrimination is bounded by the limitations of single-snapshot features, and that the results are currently limited to US highway data from 2005.

In conclusion, NOVA presents a paradigm shift from calibrating fixed models to discovering interpretable equations, revealing that human driving behavior is governed by bounded, saturating responses to relative velocity and optical looming that can be mathematically recovered from raw data.