A data-driven, model-free traffic-light controller cuts citywide travel time and CO2 in a large Zurich simulation without bespoke traffic models. The approach substitutes sensor data for costly modeling—speeding deployment and lowering operational costs—though real-world gains hinge on representative data, safety checks and online adaptation.

Main Finding

A model-free traffic control approach — combining behavioral systems theory with data-enabled predictive control (DeePC) — can steer urban traffic via dynamic traffic-light control without building explicit traffic models, and delivers measurable improvements in system-level outcomes (reduced total travel time and CO2 emissions) when validated on a very large, high-fidelity microscopic simulation of Zürich.

Key Points

• Methodological contribution: reframes urban traffic dynamics using behavioral systems theory so system evolution can be learned and predicted directly from measured input–output data (no explicit model identification).
• Control approach: applies Data-Enabled Predictive Control (DeePC) to compute traffic-light actuation sequences that respect hard constraints (e.g., signal timing and safety) while optimizing performance.
• Practical advantages: avoids the expensive, time-consuming model-building step of model-based control; better handles sparse or limited traffic measurements compared with many ML methods; enforces hard constraints explicitly (unlike many pure learning approaches).
• Large-scale validation: tested on a high-fidelity microscopic closed-loop simulation of the city of Zürich — reported as the largest such closed-loop urban-traffic simulation in the literature — demonstrating improvements in aggregate metrics (total travel time and CO2 emissions).
• Robustness considerations: method leverages measured system trajectories (Hankel/data matrices) and regularization to deal with noise and limited data; however, real-world deployment will still require representative data coverage and online adaptation.

Data & Methods

• Theoretical basis: behavioral systems theory — represents system trajectories as combinations of measured input/output data, enabling prediction and control without parametric model identification.
• Control algorithm: Data-Enabled Predictive Control (DeePC). DeePC builds predictive models from past input/output trajectories (Hankel matrices) and solves an optimization (typically quadratic or convex) over future input sequences subject to constraints and cost objectives.
• Objectives and constraints: optimize traffic performance (aggregate travel time, emissions proxies) while enforcing operational constraints on traffic signals (phasing, minimum/maximum green times, safety).
• Validation environment: a very large-scale, high-fidelity microscopic closed-loop simulator of Zürich (traffic demand, network topology, vehicle interactions modeled at microscopic level; controller is in-the-loop). Performance evaluated on system-level metrics such as total travel time and CO2 emissions.
• Comparisons: method is presented against baseline approaches (e.g., fixed-time or standard adaptive schemes), with DeePC achieving superior system-level outcomes in the simulated experiments.
• Data requirements: needs representative historical trajectories (inputs = signal settings; outputs = traffic measurements such as flows, speeds, queue lengths). Method copes with sparse sensing better than many ML baselines, but quality/coverage of data still matters.

Implications for AI Economics

• Value of data infrastructure: demonstrates how investment in sensing and data-collection infrastructure (real-time traffic sensors, connected vehicles) can substitute for costly modeling work, accelerating deployment of advanced control and increasing returns on sensor data.
• Cost savings and externalities: reductions in travel time and CO2 point to measurable economic benefits—lower fuel consumption, productivity gains, reduced pollution-related health costs—strengthening the business case for data-driven traffic control.
• Market opportunities and business models: enables scalable “control-as-a-service” offerings for cities or traffic authorities that use city data to deliver operational improvements without bespoke model development; creates a market for curated urban traffic datasets and real-time data platforms.
• Public policy and regulation: motivates policies for safe data sharing and standardization (to produce representative training trajectories) and regulation/validation frameworks to certify data-driven controllers that directly affect public infrastructure and safety.
• Distributional effects and welfare: system-level improvements do not necessarily imply uniform benefits—spatial or temporal distribution of gains may favor some routes or neighborhoods; planners must consider equity and potential unintended reallocations of congestion.
• Complementarities and rebound effects: improved throughput and lower travel costs can increase trip demand (rebound), so gains in congestion/emissions may be partially offset unless paired with demand-management measures (pricing, transit improvements).
• Labor and organizational impacts: reduces reliance on lengthy modeling efforts by traffic engineers, shifting the skill mix toward data engineering, controller integration, and ongoing monitoring; may change procurement and vendor relationships.
• Robustness and deployment risk: economic value depends on representativeness of training data and ability to adapt to structural changes (new infrastructure, demand shocks); investing in continual data collection, online adaptation, and safety verification increases deployment costs but lowers operational risk.

If you want, I can (a) produce a brief technical sketch of the DeePC optimization used here (objective, constraints, Hankel construction), (b) outline a small economic cost–benefit calculation framework to evaluate city deployment, or (c) list practical steps for a city to pilot this approach. Which would be most useful?