Comment by WalterBright
6 hours ago
There's a major arterial near my house. One day, the power went out in the area, and the traffic lights went dead. I discovered to my amusement that throughput on the arterial and its crossings increased dramatically, as drivers simply cooperated with each other.
There are already cameras on top of all the lights. How hard can it be to hook them up to AI that is instructed to maximize throughput? The current system will block 3 directions while having a green light on the 4th which has no traffic.
I once complained to a friend about this, and he told me that the traffic engineers were experts and I was lying. The next day I was driving on that road, and got a red light. You could see on down the distance 3 greens. There was no other traffic, and no cross traffic. So I sent him a photo. He conceded :-)
> How hard can it be to hook them up to AI that is instructed to maximize throughput? The current system will block 3 directions while having a green light on the 4th which has no traffic.
You don't need AI for that, you just need the system configured properly. Lights near my parents' house have been running cycles adapted to actual traffic for at least two decades, probably three... although using inductive loops in pavement rather than cameras.
There are reasons you might want a short green every so often even with no detected traffic... it would help manage worst case situations for a stuck open detector, etc. But those wasted greens shouldn't be long enough to notice. You might also have situations where locally suboptimal behaviors improves performance or safety elsewhere and that may lead to a better overall system results. I'm not claiming this is the situation in your scenario, but it can explain some negative experiences.
Red lights also regulate speed, although I don't know if they are intentionally de-optimized from green for this reason.
I used to work at a traffic signal controller company. We were under strict orders not to use AI. We did anyway, in a simulation, and let it optimize over the weekend. When we got back it had blown all of our previous attempts out of the water. We huddled around my coworker's workstation eager to see its masterful new pattern. It just made one phase red all the time and one phase green all the time, effectively closing one road. The throughput on the other road was through the roof though, since its light was always green. We did not present these results to the boss.
I guess the main reason traffic lights exist is to avoid incidents, not necessarily to maximize throughput. Maximizing throughput is just a nice no have, and so no one bothers. Just a guess though, I've no idea what I'm talking about.
A lot of lights exist not to add throughput, but to replace 2-way stops where less performant traffic (so like idiots, grandmas and heavy trucks) would have problems getting in, say fuggit (or get honked at, get flustered) and cause a near miss at least.
> There are already cameras on top of all the lights.
Those are infrared sensors that pick up the heat from car engines. They've had to be modified to pick up electric cars.
> How hard can it be to hook them up to AI that is instructed to maximize throughput?
The equipment cabinet at the intersection has been engineered to prevent all sorts of fatal things that happened in the past. Such as one set of relays being stuck so that one direction always gets a green light while the other direction gets red/yellow/green. They're designed so that can never happen again (when "broken", you'll see blinking red in all directions).
The controllers need to be standardized so that you don't need experts to be able to diagnose & repair them. They also need to be certified. It is not possible to certify an expert system (or as they're now called "AI") because it is not possible to use formal methods to prove correctness.
> The current system will block 3 directions while having a green light on the 4th which has no traffic.
That specific intersection is set on a timer. For a single, 2-lane each way, non-complicated intersection, it can cost $100k for the lights, poles, sensors and equipment cabinet. Every wire is underground in pipes so that the expansion of concrete in weather won't stretch/break the wires.
> as drivers simply cooperated with each other.
That's something that most people learn in kindergarten: "share with others" and "take turns".
Where I live (fly-over country), when the power is out, traffic backs up significantly.
Disclaimer: I used to work for my state's combination DMV & highway department.
Actual "Expert systems" could in principle get certified because they have explicable behaviour. The juice isn't worth the squeeze but it could be done. However the LLMs do not, we don't really know why this does anything even vaguely useful and we can't say if it might fail, or how we'd know. Useless.
> Those are infrared sensors that pick up the heat from car engines. They've had to be modified to pick up electric cars.
Do they pick up ICE motorcycles ? Electric motorcycles ?