![[personal profile]](https://www.dreamwidth.org/img/silk/identity/user.png){kind=small}
[Happy birthday to ![[livejournal.com profile]](https://www.dreamwidth.org/img/external/lj-userinfo.gif)
ladysprite!]
One of the more common sources of minor modern-world frustration is the traffic intersection. Traffic lights are typically pretty blunt-force instruments, running the same timing over and over again, regardless of traffic. The really fancy ones might be sensitive to time of day, and have their timing modeled in advance by traffic engineers, but it looks to me like those engineers don't come and check their work very often, because we still spend our time sitting at badly-programmed lights. And sadly, individual human ingenuity doesn't seem to help much: putting a person in charge of the intersection usually makes things worse, not better. (Nothing screws up traffic quite as effectively as the average traffic cop. A few are good at it, but most are terrible, because they rarely switch directions fast enough.)
This leads me to wonder if anybody's considered the one object that potentially has all the necessary information to do it right: the light itself. Could we build a *smarter* traffic light? I suspect so, by making a light that understands the traffic passing beneath it and the time of day, and can experiment and learn what works.
Here's the high concept. Take a traffic intersection, with all the related lights, and put sensors in each direction. (Could be treadles, could be cameras, whatever -- the important bit is that it can count the cars coming through.) Hook this whole thing to a simple learning computer -- probably a neural net, possibly some sort of annealing or evolutionary algorithm, so long as it is capable of gradually improving its own timing. Give the computer a clear measure of "better" and "worse", which mostly consists of the number of cars that actually pass through, plus a desire to balance the directions reasonably fairly.
Peg some sensible extremes: eg, don't go green for less than five seconds or more than sixty at a time. Don't allow the timing to shift too quickly, to avoid dramatically bad extremes to come out suddenly. Hard-code the firm assumptions: eg, exactly one direction must be green at any given time, and a fire truck's signal overrides everything else. Give the thing a clock (with an understanding of time and day of the week) as an additional input, so that it can factor that into its calculations.
For extra credit, add an additional sensor a bit down the street in each direction, to detect backups from the light, and put a particular priority on avoiding them.
In more complex environments, where there are a number of lights near to each other, hook the networks together, so that they can work co-operatively to improve the overall traffic flow.
For all I know, someone may already have invented this, but it sure isn't widespread if so. (This invention brought to you by my sitting in the usual pointless Route 3 traffic jam on my way home the other day.) It could make a fine project for some entrepreneurial programmer: build and improve the thing with simulated inputs, and then find an agreeable traffic-light manufacturer to partner with or sell it to. I suspect it would require a fancier computer than most lights now have, but in the age of $300 laptops, I can't imagine it would add significantly to the cost of a full light system.
This is all, BTW, very strongly based on my general approach to cognition -- when I say a "smarter" traffic light, I mean that quite literally. The above has all the elements of intelligence, albeit in a very limited domain. It has a number of heterogenenous inputs; outputs that are capable of experimenting; a neural network (preferably a multi-level one) capable of associating the outputs and the inputs in feedback loops; "instincts" that start things in the right direction and help avoid foolish extremes; and instinctive concepts of "better" and "worse" to steer it in the right direction. Far as I've been able to figure out, that's most of how intelligence works: humans seems to be mostly based on the same principles, albeit scaled up in complexity many orders of magnitude.
Do you think this would work? What clever solutions have you had for workaday problems, that should get anti-patented before some patent troll gets their hands on the idea?
ladysprite!]One of the more common sources of minor modern-world frustration is the traffic intersection. Traffic lights are typically pretty blunt-force instruments, running the same timing over and over again, regardless of traffic. The really fancy ones might be sensitive to time of day, and have their timing modeled in advance by traffic engineers, but it looks to me like those engineers don't come and check their work very often, because we still spend our time sitting at badly-programmed lights. And sadly, individual human ingenuity doesn't seem to help much: putting a person in charge of the intersection usually makes things worse, not better. (Nothing screws up traffic quite as effectively as the average traffic cop. A few are good at it, but most are terrible, because they rarely switch directions fast enough.)
This leads me to wonder if anybody's considered the one object that potentially has all the necessary information to do it right: the light itself. Could we build a *smarter* traffic light? I suspect so, by making a light that understands the traffic passing beneath it and the time of day, and can experiment and learn what works.
Here's the high concept. Take a traffic intersection, with all the related lights, and put sensors in each direction. (Could be treadles, could be cameras, whatever -- the important bit is that it can count the cars coming through.) Hook this whole thing to a simple learning computer -- probably a neural net, possibly some sort of annealing or evolutionary algorithm, so long as it is capable of gradually improving its own timing. Give the computer a clear measure of "better" and "worse", which mostly consists of the number of cars that actually pass through, plus a desire to balance the directions reasonably fairly.
Peg some sensible extremes: eg, don't go green for less than five seconds or more than sixty at a time. Don't allow the timing to shift too quickly, to avoid dramatically bad extremes to come out suddenly. Hard-code the firm assumptions: eg, exactly one direction must be green at any given time, and a fire truck's signal overrides everything else. Give the thing a clock (with an understanding of time and day of the week) as an additional input, so that it can factor that into its calculations.
For extra credit, add an additional sensor a bit down the street in each direction, to detect backups from the light, and put a particular priority on avoiding them.
In more complex environments, where there are a number of lights near to each other, hook the networks together, so that they can work co-operatively to improve the overall traffic flow.
For all I know, someone may already have invented this, but it sure isn't widespread if so. (This invention brought to you by my sitting in the usual pointless Route 3 traffic jam on my way home the other day.) It could make a fine project for some entrepreneurial programmer: build and improve the thing with simulated inputs, and then find an agreeable traffic-light manufacturer to partner with or sell it to. I suspect it would require a fancier computer than most lights now have, but in the age of $300 laptops, I can't imagine it would add significantly to the cost of a full light system.
This is all, BTW, very strongly based on my general approach to cognition -- when I say a "smarter" traffic light, I mean that quite literally. The above has all the elements of intelligence, albeit in a very limited domain. It has a number of heterogenenous inputs; outputs that are capable of experimenting; a neural network (preferably a multi-level one) capable of associating the outputs and the inputs in feedback loops; "instincts" that start things in the right direction and help avoid foolish extremes; and instinctive concepts of "better" and "worse" to steer it in the right direction. Far as I've been able to figure out, that's most of how intelligence works: humans seems to be mostly based on the same principles, albeit scaled up in complexity many orders of magnitude.
Do you think this would work? What clever solutions have you had for workaday problems, that should get anti-patented before some patent troll gets their hands on the idea?