I saw a good article on c/upliftingnews about AI improving traffic signal controllers. It’s good and all, I just can’t help but think of the “look at what they need to have a fraction of our power” meme while reading it
I saw a good article on c/upliftingnews about AI improving traffic signal controllers. It’s good and all, I just can’t help but think of the “look at what they need to have a fraction of our power” meme while reading it
Improving traffic lights doesn’t require AI, you just need sensors and some basic code to respond accordingly.
Most lights in the us run on a cycle without accounting for traffic at all. Most don’t even take into account the time of day.
Car dependent design is bad. But the us can’t even do car dependency well. You have to constantly wait at a light to leave the intersection clear for no one.
The solution is not AI the solution is having people responsible who care at least somewhat.
So I do reinforcement learning research at my university, and the coworker I sit next to everyday does traffic signal optimization using multi agent reinforcement learning and simulation. (E.g. his reseach is on stuff like this paper)
And we literally agree with you; sensors are THE problem for 90% of the inefficiency. Its rare to even know how many cars pass through in a day, or whether its 1 or 500 cars waiting at a light. However, Google knows (or can approximate), which is partially why they and they alone can get something like 30% improvement.
The other 10% inefficiemcy is coordination stuff though, which can be more difficult than you might think to fix.
outside the rush hours that is true.
in the rush hours it gets tricky because of effects like a light turning green, but traffic being jammed from a red light before. For these you need a network model and it is crazy complicated to adequately model and optimise even just a small street network.
So yeah, best solution is to reduce car traffic as a whole.
That’s why you take into account the traffic lights/intersections ahead as well. Works fine over here in NL.
what is ahead? for that you need to find out which are the main routes people take. But you also cant just give the dominant route alle the passage, because the other routes are important too. With that you get a complex network you need to optimise, where a central control uses the sensor input from the individual lights, but local contral is not sufficent.
And this is what the original comment stated, with his colleagues using reinforcement learning as one possible approach.
For a big road/street, whatever the main flow of traffic is following. So for a north-south street that’s busier than the east-west street intersecting with it, optimise the flow for traffic going north-south, including the intersections ahead. A “green wave” or “groene golf” in Dutch would work wonders. Stick to the advised speed on the digital signs and you get a wave of green lights for x amount of upcoming intersections. I’ve had them for up to 9 in a row. For the streets crossing the main road, you get some sensors, probably inductive loops to check if there are cars waiting. If there are, periodically give them green and turn the main road to red. If there are no cars on the main road (e.g., at night), you could have an extra induction loop ahead of the crossing so that the light turns green for the crossing road whenever someone approaches, before even having to stop at the light.
Sure, you could use reinforcement learning there. But you really don’t have to. Analyse the traffic for a while, and it’ll stay pretty much the same for a long, long time. Just optimise the cycles according to the time of day and day of the week and you should be good.