A group of researchers has found that AI-powered cruise control could help reduce traffic jams.
Massive traffic experiment pits machine learning against ‘phantom’ jamshttps://t.co/FMarmvwy9T
— UC Berkeley (@UCBerkeley) November 22, 2022
In a five-day experiment on the I-24 corridor near Nashville, Tennessee, USA, 100 human-driven cars participated. Each was equipped with an AI-powered cruise-control system.
The team applied two algorithms: “speed planner” and “controller.” Both use information about overall traffic conditions and the immediate surroundings to determine the optimal speed.
AI-enabled vehicles also take into account information about local traffic conditions from the I-24 MOTION corridor, where the test was conducted. The stretch is equipped with 300 sensors and 4K cameras.
“Our preliminary results indicate that even with a small share of these vehicles on the road, we can effectively alter overall traffic behavior,” said study author Alexander Bayen.
According to him, the main problem of phantom traffic jams is human driving habits.
“Driving is highly intuitive. If there is a gap ahead, you accelerate. If someone slows down, you slow down,” said Bayen.
The researcher believes that deep reinforcement learning could improve traffic flow and reduce fuel consumption.
Earlier, the results of a small UC Berkeley experiment conducted in 2016 were corroborated.
Back then, 20 human-driven cars took part in the test on a closed-loop track. Researchers noted the emergence of patterns similar to those observed on highways and busy roads.
Adding one AI-equipped car to the test reduced congestion and cut fuel consumption by 40%.
During the new trial, researchers added several new technologies. These enabled vehicles to coordinate their actions with one another, allowing them to respond to surrounding conditions.
By combining data from the cars and the road, the CIRCLES team plans to update computer simulations to better reflect the real world.
“We want to teach AI to operate vehicles in a way that is unlike humans, and also to make it socially acceptable. Throughout the test week we paid close attention to daily controller settings based on driver feedback,” said Jonathan Li, chief engineer of CIRCLES and one of the researchers.
The team hopes that, in the future, their approach will be integrated into most vehicles. Researchers are already working on scaling the technology.
However, Bayen says that, because of the enormous volume of data gathered during the trials, it could take months to obtain more precise results.
In November, Waymo taught robomobiles to create weather maps in real time to improve taxi services.
In October, Japanese researchers equipped drones with ‘eyes’.
Subscribe to ForkLog news on Telegram: ForkLog AI — all the news from the world of AI!