Red Light, Green Algorithm: The Impact Of AI-Driven Signals On Paris Traffic & Emissions

1. Introduction

With over 133,000 traffic lights and growing mobility demands, Paris's traffic management system is under increasing strain. High-density areas such as the Boulevard Périphérique, Boulevard Haussmann, and Quartier Opéra experience daily congestion, resulting in longer commute times, inefficient fuel use, and rising CO₂ emissions. The current traffic signal system, as described by Plesse (2018), relies on pre-programmed cycles that are manually adjusted by engineers. This decentralized and static approach lacks the responsiveness needed to address real-time fluctuations in traffic flow.

Advancements in artificial intelligence (AI) offer a promising alternative. AI-powered systems can dynamically adjust signal timings using live sensor and camera data, optimizing flow and reducing congestion. Studies like Pillai (2024) and Nasim et al. (2023) demonstrate how such systems can decrease idle time and improve intersection efficiency. Guo et al. (2020), through a quasi-natural experiment in China, further provide evidence that smart city initiatives, particularly AI-driven interventions, can meaningfully lessen congestion.

This study attempts to evaluate the impact of implementing AI-driven dynamic traffic light systems in Paris. By analyzing traffic flow, emissions, and congestion before and after implementation, it aims to assess whether these technologies can significantly improve urban mobility and inspire future transport policy in France and beyond.

2. Problem Statement

Paris faces severe traffic congestion, particularly in high-density areas such as the Boulevard Périphérique, major boulevards (Haussmann, Saint-Michel, Champs-Élysées), and business districts like La Défense and Quartier Opéra. The city's current traffic management system relies on pre-programmed signal cycles (Le Parisien, 2018), with adjustments made manually by engineers based on fixed scenarios throughout the day. While signal timings may be altered during peak hours (such as extending green lights or prioritizing pedestrian crossings) these modifications lack the flexibility to respond in real time to sudden congestion, accidents, or disruptions.

The system presents multiple inefficiencies. Traffic signals follow rigid schedules and are not responsive to live conditions. Manual interventions at night are slow and unable to address unexpected traffic surges. Moreover, Paris is divided into 20 independent traffic light ‘sectors,’ which operate in isolation, lacking centralized coordination. Pedestrian management further complicates traffic flow; while short wait times reduce jaywalking, they may inadvertently disrupt vehicle movement.

Despite having an estimated 133,000 traffic lights, congestion remains a persistent challenge. The Boulevard Périphérique is especially prone to bottlenecks, notably near Porte de Bercy, Porte de Saint Mandé, and Porte de Charenton, due to heavy commuter inflows. According to INRIX France, Paris ranks as the most congested city in Europe, with drivers losing 70 hours annually in traffic. The TomTom Traffic Index further reports that drivers spend 255 hours a year in transit, with 120 hours lost to congestion, and average CO₂ emissions per driver reach 1,115 kg annually. These inefficiencies not only delay commuters but also result in excessive idling, increased fuel consumption, and worsening urban pollution, signaling the need for more responsive, data-driven traffic solutions.

3. Proposed Solution

Installing AI-driven dynamic traffic lights offers a potential solution to urban congestion by utilizing real-time data from sensors and cameras to adjust signal timings. Unlike static, pre-programmed cycles, AI systems can simultaneously process live traffic volumes, public transport schedules, accident alerts, and weather conditions to deliver responsive, intersection-level coordination.

The proposed system will follow the model of SURTRAC (Scalable Urban Traffic Control), an adaptive framework developed in Pittsburgh. SURTRAC optimizes signal timings using decentralized decision-making informed by local sensor inputs, while also prioritizing public transport and adapting to environmental conditions. Applying a similar approach in Paris would enable traffic lights to dynamically respond to evolving patterns and reduce delays in high-density zones.

Beyond improving flow for daily commuters, including drivers, cyclists, and pedestrians, AI systems offer significant environmental and economic benefits. These include lower CO₂ emissions and improved fuel efficiency through reduced idling and stop-and-go movement. Time saved in traffic also translates into greater worker productivity, while more reliable public transport may encourage a shift away from car use, further relieving pressure on the network.

4. Study Design & Methodology

4.1 Research Design

This study employs a quasi-experimental design using a Difference-in-Differences (DiD) approach to estimate the causal impact of AI-driven traffic light systems on urban congestion and emissions. Intersections in Paris will be divided into treatment and control groups, with approximately 200–300 intersections in each group. To ensure representativeness, both high-traffic boulevards and quieter streets will be included, with randomization stratified by traffic volume, creating a diverse but balanced sample.

The treatment group will consist of intersections where AI traffic signal control is implemented. The control group will maintain traditional signal timing. To address potential selection bias, such as preferential installation of AI systems in already congested areas, propensity score matching will be used to ensure comparability between treatment and control intersections.

4.2 Data Collection

Data will be gathered from various sources, including traffic sensors, vehicle GPS data, pollution monitoring systems, and fuel sales records. The analysis will span a four-year period, encompassing a two-year pre-treatment phase and a two-year post-treatment phase. This allows for a comprehensive evaluation of both short-term and long-term impacts.

The study will primarily evaluate the throughput of vehicles per light cycle at each intersection, which serves as a core indicator of traffic flow efficiency and indirectly captures potential spillover effects on control areas resulting from improvements in adjacent treatment zones. To assess the broader impacts of the intervention, key outcome variables will include a comprehensive set of traffic and environmental indicators, such as average wait time per vehicle, signal cycle length, number of vehicle stops, queuing lengths, intersection throughput, average bus delay and dwell time, fuel consumption per vehicle (liters/km), estimated CO₂ and NOx emissions, and pedestrian signal delay where applicable.

Emissions will be estimated using COPERT-compatible methodologies, taking into account vehicle speed, idling time, and flow data derived from GPS and signal logs. Where feasible, stationary air quality monitoring stations will be used to validate model-based estimates, ensuring the accuracy and reliability of emission metrics. Furthermore, additional contextual data, such as weather conditions, road construction activity, public transport disruptions, and event schedules, will be included to minimize the influence of confounding factors that can affect traffic patterns.

4.3 Empirical Strategy
 

𝑌𝑖𝑡 = 𝛼 𝛽1𝐴𝐼𝑡 𝛽2𝑃𝑜𝑠𝑡𝑡 𝛽3(𝐴𝐼𝑖 ∗ 𝑃𝑜𝑠𝑡𝑡 ) 𝛾𝑋𝑖𝑡 𝛿𝑖 𝜆𝑡 𝜀𝑖𝑡

Where: 𝑌𝑖𝑡 = 𝑠𝑒𝑡 𝑜𝑓 𝑑𝑖𝑓𝑓𝑒𝑟𝑒𝑛𝑡 𝑡𝑟𝑎𝑓𝑓𝑖𝑐 𝑖𝑛𝑑𝑖𝑐𝑎𝑡𝑜𝑟𝑠 (𝑖𝑒 𝑛𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑐𝑎𝑟𝑠 𝑝 𝑒𝑟 𝑙𝑖𝑔ℎ𝑡 𝑐𝑦𝑐𝑙𝑒) =

𝐴𝐼𝑖 = 𝐷𝑢𝑚𝑚𝑦 𝑣𝑎𝑟𝑖𝑎𝑏𝑙𝑒 (1 𝑖𝑓 𝑖𝑛𝑡𝑒𝑟𝑠𝑒𝑐𝑡𝑖𝑜𝑛 ℎ𝑎𝑠 𝐴𝐼, 0 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒)

𝑃𝑜𝑠𝑡𝑡 𝐷𝑢𝑚=𝑚 𝑦 (1 𝑓𝑜𝑟 𝑝𝑜𝑠𝑡 − 𝑖𝑚𝑝𝑙𝑒𝑚𝑒𝑛𝑡𝑎𝑡𝑖𝑜𝑛 𝑝𝑒𝑟𝑖𝑜𝑑, 0 𝑜𝑡ℎ 𝑒𝑟𝑤𝑖𝑠𝑒)
𝐴𝐼𝑖 =∗ 𝑃𝑜𝑠𝑡𝑡 𝐼𝑛𝑡𝑒𝑟𝑎𝑐𝑡𝑖𝑜𝑛 𝑡𝑒𝑟𝑚 𝑚𝑒𝑎𝑠𝑢𝑟𝑖𝑛𝑔 𝑡ℎ𝑒 𝑐𝑎𝑢𝑠𝑎𝑙 𝑒𝑓𝑓𝑒𝑐𝑡
𝑋𝑖 𝑡 = 𝑉𝑒𝑐𝑡𝑜𝑟 𝑜𝑓 𝑡𝑖𝑚𝑒 − 𝑣𝑎𝑟𝑦𝑖𝑛𝑔 𝑐𝑜𝑛𝑡𝑟𝑜𝑙 𝑣𝑎𝑟𝑖𝑎𝑏𝑙𝑒𝑠 (𝑒. 𝑔., 𝑤𝑒𝑎𝑡ℎ𝑒𝑟, 𝑡𝑟𝑎𝑛𝑠𝑖𝑡 𝑠𝑡𝑟𝑖𝑘𝑒𝑠)
𝛿𝑖 = 𝐼𝑛𝑡𝑒𝑟𝑠𝑒𝑐𝑡𝑖𝑜𝑛 𝑓𝑖𝑥𝑒𝑑 𝑒 𝑓𝑓𝑒𝑐𝑡𝑠
𝜆𝑡 = 𝑇𝑖𝑚𝑒 𝑓𝑖𝑥𝑒𝑑 𝑒𝑓𝑓𝑒𝑐𝑡𝑠
𝜀𝑖𝑡 𝐸𝑟𝑟𝑜𝑟 𝑡𝑒𝑟𝑚

To ensure robustness, several strategies will be employed: parallel trends assumption will be tested using historical data; placebo tests will assign random AI implementation dates to verify that observed effects aren’t driven by unrelated trends; and spillover effects will be minimized by establishing buffer zones between treatment and control intersections. Fixed effects will also be used to control for intersection-specific and time-specific unobserved heterogeneity.



For more...