CoSyDelay: Cognitive Symbolic Regression for Traffic Delay Estimation

Lindong Xie, Xiexin Zou, Yang Zhang, Beiyu Song, Zeren Xing, Edward ChungarXiv 2026 · arXiv

TL;DR

CoSyDelay is a new framework that uses Large Language Models (LLMs) to "discover" mathematical formulas for traffic delay that are both highly accurate and physically sensible. By bridging the gap between oversimplified classical equations and opaque "black-box" machine learning, it provides traffic engineers with transparent, trustworthy tools for urban congestion management.

What Problem Does This Paper Address?

Estimating traffic delay at signalized intersections is a cornerstone of urban planning, but existing methods force a difficult choice. On one hand, analytical models (like the decades-old Webster’s model) provide clear mathematical formulas that engineers trust, but they often fail in complex, real-world traffic because they rely on idealized assumptions. On the other hand, data-driven models (like Deep Learning or XGBoost) are incredibly accurate but act as "black boxes." They offer no explanation for their predictions and can sometimes violate basic physics—for example, predicting that adding more cars will somehow decrease delay.

The authors aim to find the "Goldilocks" solution: a model that is as accurate as modern machine learning but as transparent and physically consistent as a textbook formula.

Methods

The authors introduce CoSyDelay, a "Cognitive Symbolic Regression" (SR) framework. Its workflow consists of several novel steps:

  1. Problem Decomposition: Instead of trying to find one giant formula for an entire intersection, CoSyDelay breaks the problem down into "turning movements" (e.g., the specific delay for cars turning left from the North approach).
  2. Domain-Informed Inputs: Rather than feeding raw data into the system, they transform inputs into established traffic-theory variables: Flow Ratio ($y$), Green Ratio ($\lambda$), and Cycle Length ($C$).
  3. LLM-Guided Search: Traditional symbolic regression uses genetic algorithms that often produce "mathematical alphabet soup." CoSyDelay uses an LLM (specifically GPT-4.1-mini) as a "reasoning" operator to propose and refine plausible mathematical structures (skeletons) based on traffic principles.
  4. Coefficient Calibration: Once the LLM suggests a formula structure, the system uses a standard optimizer (L-BFGS-B) to fine-tune the specific constants ($a_1, a_2, \dots$) for each movement.
  5. Composite Fitness Function: This is the "Cognitive" heart of the system. It evaluates candidate formulas not just on accuracy ($R^2$), but on seven physical principles, such as non-negativity, monotonicity (delay must increase as flow increases), and boundary behavior (delay should approach infinity if green time goes to zero).

Datasets

The study evaluated CoSyDelay across six signalized intersections:

  • Intersections 1 & 2: Simulated standard four-leg intersections used to establish a baseline.
  • Intersections 3–6: Real-world configurations based on field-collected traffic volumes and signal timings from Bowen Bridge Road in Queensland, Australia. These were modeled in simulation to generate ground-truth delay labels.
  • The final dataset contained 10,274 samples, split into training, validation, and test sets.

Key Insights and Ideas

  • LLMs as Mathematical Architects: The most striking insight is that LLMs, when prompted with domain knowledge, are better at suggesting "physically meaningful" math structures than traditional automated search methods. They tend to produce formulas that "look" like physics.
  • The "Trust" Constraint: By baking physical laws (like monotonicity) directly into the fitness function, the model is prevented from over-fitting to noise in the data. If a formula predicts that less green time leads to less delay, it is immediately discarded, no matter how well it fits the training points.
  • Shared Structure, Local Parameters: The framework discovers a single form of an equation that works across an intersection, but allows the coefficients to vary. This acknowledges that while the physics of a left turn and a through movement are the same, their specific magnitudes differ due to geometry.

Key Results

  • Superiority over Classics: CoSyDelay significantly outperformed all standard and data-calibrated versions of classical models (Webster, HCM, and Akçelik) across all six intersections.
  • Competitive with Black Boxes: It achieved accuracy levels comparable to—and sometimes better than—state-of-the-art machine learning models like XGBoost and Graph Convolutional Networks (GCN).
  • Interpretability: The model discovered a specific empirical formula (Equation 9 in the paper) that combines power-law growth for traffic demand with exponential growth for saturation.
  • Generalization: Formulas discovered at one intersection were successfully "transferred" to entirely different intersections. By simply recalibrating the coefficients, the model maintained high accuracy, proving the discovered mathematical structures are robust.

Limitations and Future Work

  • Movement Interactions: The current model treats turning movements as somewhat independent. It does not yet explicitly account for complex interactions like "queue spillback" (where a line of cars blocks an adjacent lane) or shared-lane blocking.
  • Geometric Variables: The current inputs are traffic-centric (flow, green time). Future versions could incorporate geometric data like lane widths or the length of turn pockets to further refine the coefficients.
  • Real-time Efficiency: While the resulting formulas are lightning-fast to execute, the discovery process (iterating with an LLM) is computationally heavier than simple regression. The authors suggest improving search efficiency for real-time signal optimization.

For those interested in the implementation, the authors have made their source code and datasets available on GitHub.

Summary updated

Abstract

Trustworthy estimation of traffic delays at signalized intersections is critical for modern transportation systems, as it supports proactive traffic control for congestion mitigation. Existing analytical methods often rely on idealized assumptions that limit their accuracy under complex traffic conditions, whereas data-driven methods may achieve strong estimation performance but usually lack transparent structures, which reduces confidence in their practical application. To address these limitations, we propose CoSyDelay, a cognitive symbolic regression (SR) framework for automatically deriving explicit and physically consistent empirical delay estimation expressions (DEEs). Specifically, CoSyDelay decomposes intersection-level delay estimation into turning-movement-level subproblems, thereby reducing the complexity of the symbolic search. During DEE learning, CoSyDelay constructs domain-informed symbolic inputs from traffic variables. Large language models (LLMs) are employed as symbolic search operators to iteratively generate and refine candidate expression structures, whose movement-specific coefficients are subsequently calibrated using approach-level delay labels. A composite fitness function evaluates the learned DEEs jointly in terms of estimation accuracy and compliance with specified traffic-delay principles. Experimental results on multiple signalized intersections demonstrate that the resulting expressions achieve competitive estimation performance while maintaining physical consistency, thereby supporting trustworthy delay estimation. Source code and datasets are available at https://github.com/ForrestXie9/CoSyDelay.
BibTeX
@misc{xie2026cosydelay,
  author = {Lindong Xie and Xiexin Zou and Yang Zhang and Beiyu Song and Zeren Xing and Edward Chung},
  title = {{CoSyDelay: Cognitive Symbolic Regression for Traffic Delay Estimation}},
  year = {2026},
  eprint = {2611.19946},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.19946},
}
Figure