Kill that Mocking Bot: The Myth of AI, and the Rebirth of Classical Modeling
The Siren Song of AI: A Myth for Our Time
In Homeric myth, the Sirens were enchanting creatures whose beautiful songs lured sailors to their doom. Their voices promised knowledge, insight, even ecstasy—but those who gave in were dashed against the rocks.
Today, the seductive call of artificial intelligence—particularly deep learning—can feel eerily similar. It promises to solve problems, replace experts, and reshape every industry.
In transportation, we too hear the similar voice: "Let data speak for itself." "End-to-end learning will probably render classical traffic models obsolete."
But is it true? Or are we steering blindly toward a shipwreck?
The Illusion of Bigness
Modern AI, especially deep neural network (DNN), builds (statistical data) models often described as "big" and “deep.” But what does that really mean?
Depth - Layers Upon Layers
“Deep” refers to the number of layers—dozens, hundreds, or even thousands—stacked sequentially to extract progressively abstract and higher-order feature from input data.
Parameters - A Sea of Weights and Biases
Each connection between neurons is associated with a weight (and sometimes a bias term), learned from large datasets. These parameters are purely statistical—they don’t represent physical properties, only values tuned to minimize training error.
Width - High-dimensional Feature Space
Modern DNNs operates in feature spaces with thousands of dimensions, untied to physical phenomena. This allows them to capture complex, nonlinear relationships in the data, but also makes their workings hard to interpret—even to domain experts.
Deep learning models are "big" in the scale of their parameters, and complexity of their computational structure. However, this kind of “bigness” differs fundamentally from the complexity found in classical traffic modeling—a difference that has crucial implications.
“Big” and “Deep”: the Catch
Here’s the catch: despite their scale, these models are not built on physical laws. The weights in a neural network don’t correspond to traffic flow, signal timing, or driver behavior. They’re statistical artifacts, learned solely from data. You can’t explain them, audit them, or reason causally with them.
This is why calling such models "big" is misleading. It equates size with significance. But in engineering—and especially in traffic modeling—bigness without grounding is often a liability.
Again, Structure, Not Just Size
If deep learning’s “bigness” lacks grounding, what could be the antidote?
The answer probably lies in physics-informed AI—a growing movement to embed physical laws and domain knowledge directly into AI architectures.
In physics-informed neural networks (PINNs), the model doesn’t just learn from data; it also learns to satisfy known physical equations (e.g., conservation of mass, energy, or momentum). In transportation, this could mean encoding the continuity equation, car-following behavior, or network flow constraints directly into the training process.
Physics-informed AI doesn’t ask “How many layers?” but rather “What structure does the real world impose?”
Computational Graph as an Infrastructure
A deep neural network is a computational graph—a directed acyclic graph (DAG) where each node represents a mathematical operation (e.g., non-linear activation) and each edge represents a data flow.
This graph structure not only defines how data moves through the model, but also how gradients are computed via backward propagation, and how the parameters are updated during training.
The significance of computational graphs goes far beyond just being a mechanism for representing the neurons and their connections. They offer a structured way to express complex nonlinear function approximation tasks, breaking them into layered compositions of simpler functions. This layered modeling is somewhat analogous to constructive mappings in functional spaces—gradually transforming low-level representations into more abstract feature spaces through successive combinations of linear and nonlinear operations.
From a mathematical standpoint, these hierarchical compositions in a computational graph can even be likened to how Lie groups generate global structure from local transformations in geometric spaces. Each layer’s transformation might seem small and devoid of physical meaning on its own, but stacked together, they empower the model with immense expressive power.
The Enduring Value of Classical Traffic Models
Classical traffic models are more than just equations or algorithms—they are engineering instruments, built from first principles and refined by decades of domain expertise. They reflect a long tradition of modeling grounded in physical realism and analytical rigor.
Some of the foundational approaches include:
Travel demand models: capturing long-range interactions between land use, trip generation, and modal choice.
Microscopic, mesoscopic, and macroscopic simulation models: offering multiscale representations of vehicle dynamics and network performance.
Continuum traffic flow model: a partial differential equation modeling kinematic waves in traffic flow.
Network equilibrium models: solved with convex optimization techniques like Frank-Wolfe to represent user behavior and system performance under congestion.
These models are not black boxes. They are transparent, interpretable, and often causal. Variables have clear definitions—speed, flow, density, demand, delay. Parameters correspond to physical or behavioral quantities—headways, saturation flow rates, travel time elasticities. And most importantly, these models are governed by domain-specific physics and analytical abstractions, not just statistical fitting.
Their structure is not arbitrary—it mirrors the structure of the real world.
True Complexity Comes from Physical Reality
The complexity of classical models isn’t manufactured—it’s inherited from the real world.
Unlike deep neural networks, where complexity grows with architectural choices (more layers, more parameters), the “bigness” of classical models emerges organically from the systems they represent. Their structure reflects the real physical and institutional landscape of transportation:
Large-scale networks with intricate geometries and interdependencies
Diverse traveler behaviors, choices, and constraints
Random variations in demand, incidents, and control strategies
Dynamic feedback loops from pricing, policies, infrastructure, and human adaptation
In this sense, classical models scale with the problem, not with the size of a neural net. Their complexity mirrors the underlying system—not an abstract optimization objective.
By contrast, deep neural networks create complexity through design, not discovery. Their depth and parameter count are the results of tuning and architecture choices, not reflections of how the real world works.
Each additional layer and neuron makes the model better fit the data, but not necessarily more truthful. Their internal logic is statistical, not physical—high-dimensional and expressive, but often untethered from the causal dynamics of traffic flow, signal timing, or traveler behavior.
Let’s keep in mind: classical models grow in complexity because the world is complex. DNNs grow in complexity because—we made them that way.
Integration: A Path Forward
Classical traffic models are engineering tools. Deep neural networks are data-driven approximators. The strength of classical models lies in their physical consistency and transparent decision-making, these give classical models strong interpretability. They can answer causal questions like "What happens to congestion if we add a lane?"—making them central to causal inference and policy analysis.
For evaluating strategies like congestion pricing or traffic signal control, classical models remain the gold standard.
Deep Learning Frameworks + Classical Models
The future of traffic modeling may lie in the fusion of classical methods and deep learning computational frameworks.
Yes, not just deep models, but the frameworks they provide—computation graphs, automatic differentiation, parallelization, and GPU acceleration.
Imagine:
Embedding microscopic, mesoscopic, and macroscopic traffic models inside differentiable architectures.
Using prior knowledge from traffic flow theory to improve explainability and robustness (i.e., XAI).
Feeding back domain-specific constraints into the training process to mitigate exploding/vanishing gradients.
Designing attention mechanisms tailored for traffic systems to avoid pathological lost surfaces, such as flat regions, sharp minima, and saddle-point—by embedding domain-specific insights directly into the model architecture .
The integration can further include:
Crafting loss functions that respect traffic flow constraints (e.g., conservation laws, capacity limits)
Designing activation functions that preserve gradient flow over temporally sparse yet semantically dense events (like signal phase changes or shockwave formations).
Such integration can smooth the optimization landscape, mitigate instability during training, and improve generalization—something generic data-driven architectures could fail to capture in structured domains like traffic systems. This hybrid paradigm preserves the causality of physical models while gaining the expressiveness of statistical learning.
Toward Domain Specific Large Model
A true “traffic-domain-specific large model” doesn’t have to be one that imitates GPT or taking the Large Language Model (LLM) route. Instead, it could:
Combine physics-based traffic theory with neural computation
Preserve interpretability and policy relevance
Leverage AI to handle perception and pattern recognition tasks—like video interpretation and trajectory prediction (See Table 1 below)
Use engineering insights to inform model structure, optimization, and robustness
Kill that Mocking Bot, or Not?
Probably we don’t need to kill the mocking bot. We just need to teach it to harmonize with the deep, steady voice of classical modeling.
Deep neural networks offer powerful computational tools—in traffic modelling domain, their true potential could be maximized when they are grounded in the solid principles of classical modeling—models that are interpretable, causally driven, and rooted in the physical reality of traffic systems.
AI isn’t the enemy of classical models—it’s a complement, an enhancer, and a partner. By informing computational models to reflect the structure and dynamics of the real world, we can create systems that are not just "big" in terms of data or layers, but meaningful in their ability to understand and predict the flow of people, goods, and vehicles.
This is the path forward: a harmonious blend of engineering rigor and data-driven innovation that will ultimately drive more accurate, efficient, and adaptable traffic modeling paradigms.
👉 👉 This article has a Chinese Version.