How AI Turns Traffic Lights Into Intelligent Agents

As reported by ReadWrite: Today’s new cars are loaded with sensors and powerful computer processors. That’s the high-tech pathway to turning our vehicles into super-efficient, semi-autonomous—or even self-driving—"transportation devices."

Unfortunately, the roads these clever mobility machines drive on are all too often, well, dumb. You experience the pain of this problem every time you senselessly wait for an extra couple minutes at a red light, when there are no other cars in sight from any direction.

Samah el-Tantawy, a recently minted PhD of Engineering from the University of Toronto, wants to change that.

Inspired by research from her advisor and director of the Toronto Intelligent Transportation Systems Centre, Professor Baher Abdulhai, el-Tantawy devised a system that uses artificial intelligence and game theory that, in a simulated environment, shaved 40% of the time off an average wait at an intersection. She said that could mean 12.5 fewer minutes stuck in your car, if you pass through about 30 intersections on your commute.

Can We Talk?

According to el-Tantawy, many of today’s traffic lights at intersections operate based on pre-programmed repeated cycles that run with little or no input from fluctuations in traffic. Yes, there are sensors in pavements along major arteries, but those inputs into centralized systems might only be able to extend a green light for a few seconds. Like other centralized disconnected top-down systems, there are inherent limitations.

Instead, el-Tantawy’s system—dubbed MARLIN for Multi-agent Reinforcement Learning for Integrated Network (of Adaptive Traffic Signal Controllers)—uses video cameras, other vehicle data inputs (if available), processing power, and routers to analyze how many drivers are zipping through the intersection and how many are simmering with road rage for wasting countless minutes at a red light. With MARLIN, cameras are aimed at all four approaches, and the system is distributed throughout a region rather than just on main streets.

“Our approach is decentralized, where the intelligence or math to assign the greens is done on the fly at each intersection,” she told me.  “The brain sits at each intersection, and calculates the best timing to minimize the number of cars approaching and waiting, and it coordinates those decisions with other lights at other intersections.”

The Shortest Wait Wins

El-Tantawy said no amount of math can perfectly model every situation. There are too many variables. The solution? “Each intersection is connected to the neighboring or adjacent intersection, sending and receiving information about the waiting vehicles,” she said.  Then, “reinforcement learning” comes into play.

Like a child learning to walk by making minute adjustments, each traffic light—or “agent,” as el-Tantawy calls them—makes a decision every second about the best way to keep motorists and pedestrians waiting for as short a period as possible.


“The agents learn, until they converge, with each one getting the best response action to achieve its goals, without negatively affecting the others. We use multi-agent reinforcement learning,” she said. “And it cascades throughout the system. The decisions by agents affect each other, so it’s a game.”

The system has to be simulated in a test environment before being placed in the street, where the learning can continue in real-world conditions. So far, MARLIN has only been used in a test environment—but with great results. That encouraged el-Tantawy and Professor Abdulhai to recently form a start-up traffic tech company to commercialize the system, and get it on as many streets as possible.

It's easy to see how this AI-powered traffic light system could be a major boost to a region's productivity, justifying the anticipated cost of about $20,000 to $40,000 per intersection.

el-Tantawy said her start-up will soon sign up its first municipality to run a field test, but wasn’t ready yet to disclose the location. I hope it’s in my neighborhood.