Bottlenecks have the well known phenomenon of capacity drop, visible below, where increasing inflows to a system lead to a sudden decrease in outflow. Normally, this is avoided using traffic light ramp metering, but you can use autonomous vehicles instead! Unlike a traffic light, which can only control the inflow at one point, the autonomous vehicles can re-adjust their velocities to enable optimal merging. We call this control technique Lagrangian Control as the autonomous vehicles can apply control at any point in the traffic stream. Using deep reinforcement learning and Flow (https://flow-project.github.io) we can learn effective controllers that can help decongest this bottleneck.
Inflow vs. outflow curve for a 4 lane bottleneck. Note the drop off after 1400 vehicles per hour.
In the video, you can see the congestion quickly form due to the inflow of 2000 vehs per hour
Human baseline - no AVs
Controlled, AVs in red