Researchers at Chung Ang University in South Korea have developed an algorithm to improve traffic flow at traffic lights.
Normally, the AI in the car is trained to recognize the traffic light or the traffic light is networked with the vehicles. But the approach is new and based on an AI in the Traffic light. The so-called meta-reinforcement in the field of machine learning for traffic lights streamlines traffic and prevents congestion.
Researchers at South Korea’s Chung Ang University use reinforcement learning (RL) algorithms to control non-stationary traffic lights. The reinforcement reward is based on the classification of traffic regimes, which is done automatically. The learning process takes place with regard to the agent (traffic light) using the try’n’error principle. This is how the AI is trained, if possible smooth traffic to allow. The AI needs to look at the big picture when it comes to vehicle deceleration and not individual vehicles.
Conventional traffic lights with fixed green and red phases are not good for improving the congestion situation and of course they cannot adapt to changing traffic. While RL algorithms can be a solution, they are not always the most effective solution in dynamic environments. The Meta-RL model is intended to remedy this.
The two main goals of this model are traffic flow and minimizing waiting times during rush hours. This is achieved with the Extended Deep Q-Network (EDQN)-incorporated context-based meta-RL model. To do this, it defines traffic as “saturated” or “unsaturated” using a latent variable related to the overall status. Because of the river, vehicle passage is increased or delays minimized. It implements cycles that are controlled by “rewards”. The EDQN as a decorder is used to control several crossings.
The tests took place in the southwest of Seoul in the area of 15 intersections, which the AI adapted to without changing any parameters – including changes in traffic conditions. Thus it had become the more effective.