Hacker News Reader

Comment by WalterBright | Hacker News Reader

> This has been done for quite some time, and it's unclear what value deep learning would add beyond the arithmetic we've used for a while.

I've seen a lot of these claims, and they're easily debunked by driving through a city. The usual case is a platoon of 5 cars cruising down an arterial. A solitary car pulls up on a side street, and the light turns read for the arterial just in time to stop the platoon for the solitary car. A sensible algorithm would let the platoon pass first.

There are endless scenarios of these, to the point where the light algorithm finds a local optima where it is completely out of phase with the traffic and causes maximum delay and minimum throughput.

A relatively small amount of effort going into cameras and an adaptive system that minimizes a cost function would save billions of dollars in gas, time, and unneeded road widening.

View on HN

There are two scenarios that inevitably get me thinking about smarter traffic lights. The first, is what you described, and the second is when I notice myself and a few others are sitting at a stop light while the intersection remains unused for 30 seconds.

> A sensible algorithm would let the platoon pass first.

You're taking an example of a sub-optimal local solution and claiming that is proof that the global solution is sub-optimal. That does not follow.

It could easily be, taking your example, that the platoon of 5 cars must stop at some future light, and the delay for 1 car is the smallest delay available for those cars. Therefore, stopping for the 1 car reduces the platoon's total travel time. As the driver, it's easy to think that was dumb because you'll never see the 15-car wait that was avoided by stopping early for 1 car.

Essentially, you are claiming that a greedy algorithm for traffic would always find the globally optimal solution (which we know is not true).

Walter isn't arguing for a greedy algorithm. He's giving an example of a local inefficiency and a mitigation. Lots of local inefficiencies add up to a global inefficiency. Illustrating a global inefficiency in itself isn't feasible in an anecdote. Thus I think your argument by assertion isn't possible to refute: you can bring up "local inefficiencies don't necessarily mean global inefficiencies" to dismiss all arguments that the current system is less than optimal.

Me, I've been stuck at traffic lights for 5 minutes at 2 in the morning, with no traffic whatsoever around. A very local algorithm would be highly preferable at that time of night, assuming usual (i.e. very little) traffic at that hour. It's an extreme example, yes: but there does exist a range of traffic patterns between the small hours of the morning and rush hour, and the same amount of global optimization effort is not applied to all hours, for obvious reasons. Something that could analyse the traffic using cameras and sensors and optimize wouldn't have to be paid extra to work late at night.

Several cities take the even simpler approach of having lights that change from timed cycles to blinking red/yellow after a certain hour.

Cars stopping and then starting again costs time, money, fuel, wear & tear, pollution, etc. It's far from an efficient solution.

Might not apply where you are, but in my state, if you sit at a red light for that length of time, with no traffic coming the other way, it is legal to ignore the light and go anyway - basically treating the light like a four-way stop sign.

Anecdotal but London seems to have a reasonably 'smart' traffic light system. It's something you only really notice when you're looking out for it (sign of a good system?), but as a pedestrian it's very often that pedestrian crossing phases match extended gaps in traffic flow. On roads with a good sightline, I often notice a gap in the flow ~30 seconds away that the lights will change in sync with.

The system is called SCOOT if you are interested.

Santa Clara, CA trumps out "retime the traffic lights" every time they want to raise taxes. Biggest crock of B.S. I've seen in a ballot.

Reply on news.ycombinator.com