Smart Roundabout based on Deep Reinforcement Learning

Industry

Services

Technology

The intersection of two elevated high-traffic tracks, solved by an unharmed roundabout, causes traffic jams of hundreds of meters and more than 6 minutes of maximum waiting time.  Because of the roundabout preference system, it can happen that the inlet flow through a high-traffic branch is interrupted by vehicles from another branch with less intensity, which has priority because it is on its left. To solve cases like this it is often necessary to improve the infrastructure, for example by building an intersection at different level, which is an important economic investment. Sometimes traffic can be improved through a Smart Roundaboutwhich learns how to regulate traffic through Deep Reinforcement Learning.

Deep Reinforcement Learning

To train an intelligent system like this, we developed a roundabout model  capable of simulating its actual behavior in different traffic situations. The traffic lights at the roundabout are then modeled and an Agent is put to control the traffic light phases. The agent receives the queue  length information  on each branch and  decides which traffic lights should turn red at any given time. From time to time, you receive a reward if you manage to reduce the maximum waiting time of the roundabout.

At first, the Agent manipulates the phones randomly, and receives penalties for it. Over time, learn the optimal way to operate traffic lights at all times, depending on the traffic situation. This agent, trained in the roundabout model, will be able to improve traffic at the actual roundabout, and continue learning.

Benefits

  • Improved maximum waiting time at  roundabout
  • Economic solution to new infrastructure implementation alternatives
  • Possibility to implant the system in other roundabouts

Indicators

The longest rush hour waiting time is reduced from 440 seconds to 285 seconds, by 35%

DOWNLOAD THE PDF

More cases

Process Partner: Anomaly Detection in manufacturing
Smart Roundabout based on Deep Reinforcement Learning
Purchase Order Reader: Automated Data Extraction