Data-driven generalized perimeter control: Zürich case study

This paper proposes a data-driven generalized perimeter control method using behavioral systems theory and data-enabled predictive control to optimize urban traffic via dynamic traffic lights, validated through a high-fidelity microscopic simulation of Zurich that demonstrates improvements in total travel time and CO2 emissions.

The Gist

Imagine a city like Zürich as a giant, living organism. Its roads are the veins, the cars are the blood cells, and the traffic lights are the heart valves. When everything flows smoothly, the city is healthy. But when too many cars try to squeeze through at once, the "blood" clots, causing a traffic jam (congestion). This clog wastes time, burns extra fuel, and pollutes the air.

For decades, city planners have tried to fix this by building a perfect mathematical map of how traffic moves. They tried to write a giant recipe book that predicts exactly how every car will behave. But cities are messy, chaotic, and constantly changing. Building that perfect recipe book is like trying to write a manual for a storm while standing inside it—it's expensive, slow, and often wrong.

The Big Idea: "Learning by Doing" instead of "Studying the Manual"

This paper introduces a new way to control traffic lights. Instead of trying to build a perfect theoretical model of the city, the authors use a method called Data-enabled Predictive Control (DeePC).

Think of it this way:

• The Old Way (Model-Based): You try to learn how to ride a bike by reading a 500-page physics textbook about balance, friction, and aerodynamics. You spend years studying, but when you finally get on the bike, you might still fall over because the real world is different from the book.
• The New Way (DeePC): You just get on the bike. You look at where you've been, feel how the bike leans, and adjust your balance in real-time. You don't need to know the physics equations; you just need to know what worked in the last few seconds.

How It Works: The "Time-Traveling" Traffic Light

The researchers treat the city's traffic data like a giant library of past experiences.

1. The Library: They feed the computer years of traffic data (how many cars were here, how fast they were going, what the lights did).
2. The Pattern: The computer looks for patterns in this library. It doesn't care why the traffic moved that way; it just knows that "When we did X at the lights, the traffic density became Y."
3. The Prediction: When a new traffic jam starts forming, the computer asks, "Based on everything we've seen before, what is the best thing to do right now to keep things moving?"
4. The Action: It adjusts the traffic lights (specifically, how long they stay green) to steer the traffic away from a "gridlock" state.

The "Zürich" Test Drive

To prove this works, the team didn't just test it on a small grid of 4 streets. They built a digital twin of the entire city of Zürich. This is a super-realistic video game simulation with:

• 15,000 roads
• 7,000 intersections
• 170,000 cars (simulating a busy evening rush hour)

They compared three scenarios:

1. The Baseline: The current traffic lights, which just follow a fixed, boring schedule (like a metronome that never speeds up or slows down).
2. The Old School AI (MPC): A smart system that tries to use a simplified math model to predict traffic.
3. The New AI (DeePC): The "learn by doing" system described above.

The Results: A Smoother Ride

The results were impressive. The new system (DeePC) was the clear winner:

• Less Time Wasted: Drivers spent about 18% less time sitting in traffic compared to the baseline.
• Cleaner Air: Because cars weren't idling and stop-and-go driving as much, CO2 emissions dropped by about 10%.
• More Trips Completed: More people actually reached their destinations during the simulation.

The Secret Sauce: The "Snake" and the "Hankel Matrix"

The paper uses some fancy math terms, but here is the simple version:

• The Snake: To manage a city this big, they can't look at every single street individually. They used a "snake algorithm" to group the city into neighborhoods (regions) that behave similarly. It's like grouping a choir by voice type (sopranos, tenors) rather than trying to conduct every single singer individually.
• The Hankel Matrix: This is the computer's "memory bank." It's a giant grid that stores past traffic patterns. The magic is that this grid can predict the future without needing to know the complex physics of cars. It essentially says, "We've seen this pattern before, and here's how we fixed it."

Why This Matters

The biggest breakthrough is that this method doesn't need to know where the traffic lights are placed relative to the neighborhoods.

• Old Method: You had to put traffic lights exactly on the border between two neighborhoods to control the flow. If the lights were in the middle of a neighborhood, the math broke.
• New Method: The computer figures out the connection itself. It's like a conductor who can hear the whole orchestra and tell the violin in the back row to play louder, even if they aren't sitting next to the violins in the front.

In a Nutshell

This paper shows that we don't need to be geniuses at traffic physics to fix traffic jams. We just need to listen to the data. By letting the traffic lights "learn" from the city's own history, we can make our cities flow smoother, save time, and breathe cleaner air. It's a shift from trying to control the city with a rigid rulebook to guiding it with a flexible, intelligent intuition.

Technical Summary

1. Problem Statement

Urban traffic congestion poses significant challenges regarding travel time, economic loss, and environmental impact (CO2 emissions). Traditional model-based control approaches, such as Model Predictive Control (MPC), rely on explicit mathematical models of traffic dynamics. However, modeling large-scale urban networks is expensive, time-consuming, and often inaccurate due to the system's complexity, stochastic demand, and dynamic infrastructure changes.

Conversely, machine learning approaches (e.g., Deep Reinforcement Learning) often require vast amounts of training data, lack theoretical guarantees, and struggle with sparse traffic data and hard constraints. There is a critical gap between theoretical research and practical implementation, largely due to the lack of common, high-fidelity benchmarks that reflect real-world complexity.

The paper addresses the Intersection Traffic Signal Control Problem (ITSCP) by proposing a Data-enabled Predictive Control (DeePC) framework. The goal is to optimize traffic flow and reduce emissions without requiring an explicit parametric model of the traffic system, using only historical data to drive control decisions.

2. Methodology

A. Behavioral System Theory Formulation

The authors model urban traffic dynamics using Behavioral System Theory. Instead of identifying system parameters (like matrices $A, B, C, D$), they define the system by its "behavior"—the set of all possible trajectories compatible with the system.

• Inputs ($u$): A combination of controllable inputs (traffic light split ratios, $\lambda$) and uncontrollable inputs (exogenous demand, $d$).
• Outputs ($y$): Traffic density ($\rho$) in specific regions of the city.
• Key Insight: The authors conjecture a linear relationship between the traffic light control input ($\lambda$) and the resulting traffic density ($\rho$), despite the underlying non-linear relationship between density and flow (described by the Macroscopic Fundamental Diagram, MFD).

B. Data-enabled Predictive Control (DeePC)

The control algorithm utilizes the Fundamental Lemma of behavioral systems. It constructs a Hankel matrix from offline recorded input-output data to represent the system's dynamics non-parametrically.

• Optimization: At each time step, DeePC solves a quadratic program to find the optimal future control sequence. The cost function minimizes the deviation of the predicted traffic density from a reference trajectory (the critical density, $\rho_{cr}$, derived from the MFD) while penalizing control effort.
• Constraints: The algorithm enforces physical constraints (e.g., density must be non-negative and below gridlock limits) and operational constraints (e.g., traffic light cycle times).
• Generalization: Unlike classical perimeter control which assumes actuators are at region boundaries, this formulation allows actuators (traffic lights) to be placed anywhere in the network. The Hankel matrix implicitly maps these distributed actuators to the aggregate network dynamics.

C. Simulation Framework

To validate the method, the authors developed a high-fidelity, closed-loop microscopic simulation of Zürich, Switzerland, using the SUMO (Simulation of Urban MObility) package.

• Scale: The network includes ~15,000 roads, ~7,000 intersections, and serves over 170,000 vehicles during peak hours.
• Benchmarking: The study compares DeePC against:
  1. Baseline: Static, pre-timed traffic signals.
  2. MPC: A state-of-the-art linearized MPC formulation that requires explicit modeling of the MFD and linearization of traffic dynamics.
• Region Partitioning: The city is partitioned into homogeneous regions using a parallelized "snake clustering" algorithm based on real traffic density data, creating a Macroscopic Fundamental Diagram (MFD) for each region.

3. Key Contributions

1. Flexible Behavioral Formulation: The paper introduces a generalized perimeter control framework based on behavioral theory that seamlessly integrates different data types and actuator locations, removing the need for actuators to be strictly at region boundaries.
2. Linear Conjecture Validation: The authors hypothesize and validate via extensive simulation that the relationship between traffic light split ratios and traffic density is approximately linear, allowing DeePC to bypass the non-linear complexity of flow-density relationships.
3. Large-Scale Benchmark: The authors present what is, to their knowledge, the largest closed-loop microscopic simulation in traffic control literature (Zürich city-scale) equipped with demand estimated from real data. They propose this simulation as a standard benchmark for future research.

4. Results

The performance was evaluated on two networks: a lattice network (64 intersections) and the full Zürich network.

Lattice Network Results:

• Gridlock Prevention: The baseline simulation resulted in gridlock during peak demand, with only 4,636 vehicles completing trips. Both DeePC and MPC prevented gridlock.
• Performance: DeePC outperformed MPC in all metrics (travel time, waiting time, and emissions), attributed to DeePC's implicit model capturing dynamics better than the linearized MPC model.

Zürich Network Results (Evening Peak):

• Travel Time: DeePC reduced average travel time by 18.08% compared to the baseline and 1.4% compared to MPC.
• Emissions: DeePC reduced CO2 emissions by 16.17% compared to the baseline.
• Trip Completion: The number of completed trips increased by 0.74% over the baseline.
• Control Period Sensitivity: The study tested control updates every 1, 3, and 6 duty cycles. While performance degraded slightly with longer intervals, DeePC remained superior to the baseline and competitive with MPC even with less frequent updates.
• Input Analysis: Principal Component Analysis (PCA) of the control inputs revealed that the first principal component explained ~48% of the variance, indicating a strong global pattern where the controller adjusts green times at city entry points (positive correlation) versus city center (negative correlation) to manage congestion.

5. Significance and Conclusion

This paper demonstrates that Data-enabled Predictive Control (DeePC) is a viable and superior alternative to traditional model-based control for large-scale urban traffic management.

• Elimination of Modeling Burden: The method replaces complex, error-prone modeling with a data collection step, making it scalable to cities where detailed models are unavailable.
• Robustness: By relying on data trajectories rather than a specific parametric model, the controller adapts to the actual system behavior, including non-linearities that linearized MPC might miss.
• Practical Applicability: The successful application to a real-world city (Zürich) with realistic demand and infrastructure proves the method's readiness for real-world deployment.
• Future Directions: The authors note that while the method is powerful, ensuring robustness under varying exogenous conditions (demand shifts) and addressing scalability for extremely large numbers of actuators (via distributed optimization) are key areas for future research.

In summary, the study establishes a new paradigm for traffic control where data-driven, model-free strategies can effectively manage complex, non-linear urban systems, significantly improving mobility and sustainability metrics.