Anyone can start a traffic jam—all it takes is tapping on your brakes. The driver behind you will brake, as will the next driver, starting a shockwave of stop-and-go reactions that can travel backward for kilometers. Now, scientists have shown that a few self-driving cars can prevent such jams—and in some cases double the average speed of surrounding vehicles.
The researchers used a video-game style interface to control simulated cars on made-up roadways. In one scenario, the cars drove around a figure eight with a central intersection. In others, one or several lanes of traffic merged, or the cars traversed a Manhattanlike city grid with traffic lights at each crossing. The team looked at various ratios of self-driving cars mixed with regular cars that simulated typical human driving.
In each scenario, the researchers tested four algorithms that used reinforcement learning—a type of artificial intelligence (AI) that learns skills through trial and error. In the figure eight and merging scenarios, a central algorithm controlled all self-driving cars, experimenting by changing their patterns of acceleration and braking. In the Manhattan scenario, the AI-controlled traffic lights instead of cars.
The results were impressive. In the figure eight scenario, replacing just one of the 14 “human”-driven cars with a self-driving car doubled the average car speed, the researchers reported last month at the Conference on Robot Learning in Zurich, Switzerland. In the merge scenarios, replacing 10% of the regular cars with self-driving cars also increased overall traffic flow, in some cases doubling the average car speed. The self-driving cars sped up traffic in part by keeping a buffer between themselves and the cars in front of them, forcing them to brake less often. Giving the algorithm control over traffic lights in a Manhattan-style traffic grid increased the number of cars passing through by 7%.
The tested algorithms leave plenty of room for improvement, says study author Eugene Vinitsky, an artificial intelligence researcher at the University of California (UC), Berkeley. That’s why his team is making its programs public. “If anyone has a brilliant solution or algorithm, you can use this framework to test [new ideas],” says Meng Wang, a transportation engineer at the Delft University of Technology in the Netherlands who’s done related work.
Researchers in other areas have created benchmarks for reinforcement learning, and “it’s great that they’re doing it in traffic,” says Daniel Lazar, an electrical engineer at UC Santa Barbara. “I hope to see the work expanded” to control not just car speed but also lane changes.
Vinitsky can’t predict when real self-driving cars will help us all get to work faster, but he says some of the new techniques could help our current vehicles. For example, the patterns of traffic-reducing acceleration and braking could be used by the adaptive cruise control systems common in new cars, he says, saving time, gas, and possibly lives. “All the tools are there.”
Source: Science Mag