The ACC/Midpoint controllers cause jams
cathywu opened this issue · comments
Example:
This is the first positive result, achieved by reducing the follow_sec to 1.0 (from 3.0):
Params: follow_sec = 1.0, max_speed = 40, gain = 0.1, beta = 0.9.
With only a few robot cars, on the other hand, (and a simple policy that means that they can't change lanes) they can have very little effect. I think this is due to the fact that they are not in the right place to make a difference.
With 50% penetration, the robot cars (using ACC) are able to improve speed:
ACC first results!
1 lane, 25 agents, all robotic agents (ACC "works")
Note: ACC still kicks in halfway through the simulation.
1 lane, 25 agents, half robots (ACC improves things)
2 lanes, 25 agents per lane, half humans (slow)
Still unsure why the changeFasterLane seems to result in super slow vehicles overall, see #28.
2 lanes, 25 agents per lane, half robots (ACC causes jams)
On the other hand, the heuristic controller (currently configured to keep a headway of 0.25 the distance between the front and back vehicles) seems to improve things!
Parameters: max_speed=40, gain=0.1, beta=0.9, duration=250, bias=1.0, ratio=0.25.
WARNING: Everything is pretty finicky right now and requires manual parameter selection and fine tuning.
I've run this a bunch of times. Every few times, the plot looks like the controller does cause traffic, but most of the time the speed looks improved.