Can AI control city wide traffic lights to reduce traffic pressure or waiting times ?

What does it mean to let AI take the reins on a city’s traffic lights? In essence, it’s about using algorithms to constantly adjust signal timings in real time, aiming to smooth out traffic flow and shrink wait times at intersections. The promise is a quieter city, less gridlock, and quicker routes. But how far has this idea actually traveled from the lab to the streets?

Background

AI-driven traffic-light control systems have moved from pilot trials to full deployments in several urban centers. These deployments rely on live feeds from intersection cameras, inductive-loop sensors embedded in roadways, and data uploaded by connected vehicles to infer current and impending traffic conditions (Nature, 2023). Machine-learning models—often trained on historical signal logs and incident reports—forecast short-term demand; reinforcement-learning agents then translate those forecasts into signal-phase decisions that minimize cumulative vehicle delay and queue lengths.

Early academic work dates back to the late 2000s, when researchers at Carnegie Mellon and the University of Texas demonstrated adaptive traffic controllers that outperformed fixed-time plans by 15–20 % during peak hours. By the mid-2010s, systems such as SCOOT (Split, Cycle and Offset Optimization Technique) and SCATS (Sydney Coordinated Adaptive Traffic System) had already been running for decades, but their closed-loop optimizations were typically heuristic rather than learning-based. The 2016 launch of Pittsburgh’s “SURTRAC” system marked the first large-scale reinforcement-learning deployment: edge devices at individual intersections learned local policies that were later coordinated by a central scheduler, cutting travel times on key arterials by roughly 25 % in field tests.

Subsequent deployments broadened both scope and technique. In Hangzhou, China, an AI engine named “City Brain” ingests feeds from 5,000 cameras and adjusts 12,000 signals city-wide, achieving a reported 10 % reduction in average trip duration. Singapore’s Green Link Determining (GLIDE) adaptive system, introduced in 2019, uses vehicle re-identification and queue-length estimation to shift green-time allocation in real time, yielding a 12 % drop in congested-peak delays. In the United States, the Federal Highway Administration’s “AI for Traffic Management” initiative has seeded adaptive algorithms in Austin, Pittsburgh, and Los Angeles, where early results show queue lengths shrinking by 18–22 % on instrumented corridors.

Beyond reducing delay, these systems aim to lower emissions by cutting stop-and-go cycles. A 2021 simulation study published in Transportation Research Part D estimated that city-wide adaptive control could cut CO₂ emissions by roughly 5 % and NOₓ by 7 % across a mid-sized metropolitan network. Emergency-vehicle pre-emption—first trialed in Kansas City in 2018—further bolsters safety metrics by granting light priority while preserving green splits for conflicting phases.

Still, open challenges remain. Data-quality issues—missing sensor feeds, camera occlusions, and adversarial spoofing—can degrade model performance. Intersection-level policies must be harmonized across districts to avoid gridlock migration; co-learning with connected vehicles promises to mitigate this by providing richer upstream demand information. Privacy and cyber-security concerns have prompted cities to adopt federated learning architectures where raw video never leaves local edge nodes. Economic barriers, especially in low-income municipalities, persist: hardware retrofits can exceed US$2,500 per signal head, though cloud-based controller-as-a-service models are beginning to lower entry costs.