Numenta 2015 Spring NuPIC Hackathon, 12 Jun 2015

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Trafficwerks [DEMO #11] (2015 Spring NuPIC Hackathon)

Description

Using NuPIC to get live traffic anomalies.

http://nupic2015spring.challengepost.com/submissions/37836-traffikwerks

A

Alright, so this is traffic works, I'm Sean. This is Tom and we're interested in climate change. So we see everyone just got these flash flood warnings right. So I'm not going to say one flash flood does not mean climate change, but we are getting a lot of weird rather recently go the next slide please. So this is where I used to live. Austin Texas, a lot of bed flubbing flooding in Houston as well.

A

I have a friend that just had the posted pictures on facebook that his whole house is just down and there's, like hundreds of people, are out and this this again, this one anomaly does not show climate change but we're seeing all these flooding on the west on the east coast and we're seeing drought in the west coast. And so what can we do about this thing? Well, we know that greenhouse gases are somewhat to do with climate. Changing and one put one major contributor of greenhouse classic gases is transportation now not a lot.

A

We know that it's probably about twenty five percent of in New York, all the transportation, all the greenhouse gases, but anything we can do to decrease the how carbon is going into the atmosphere is something that we can be. We can be proud of. So what we wanted to do is take transportation information from the NYC DLT and bring it into new pic and see.

A

Can we predict what's going on the traffic and if we can predict and see anomalies in the traffic system, then maybe we can re-route traffic to go around where places where there's um stops. If traffic is blocked up, then that makes cars are idling and having more on greenhouse gas emissions, so Google Maps does a pretty decent job of this right.

A

Now, if you go to google maps, you can see right now that there's red lines place, and that means things are blocked up, but which is great except that rush hour, oh blocked up, and so it doesn't really give you a lot of information to know that things are blocked up right now, or rather it's better to know what is really different from normal and that's what we think new pic can do.

A

We can see over the long period of time when prepped when traffic is worse than normal and then and only then give notifications and say hey, you should go somewhere else. So that's the idea we didn't get that far, but we do. We did get some information, so we went to the NYC.

A

Do t went to their stream of data and you'd nice it'd be nice if they had a nice API that we could pull data from well, what they actually have is a CSV file, which would be nice if this was actually queryable, but this is real time data, so any is no historical data.

A

So what we've been doing to get data is for the past when, when did you say, hey just start grabbing cache data right now yesterday morning, so we've been cashing data yesterday morning, every minute pulling all this data stashing the system and feeding it into new pic, let's see anything else about that data. So what we've got here is speed. Data and we've got around different paths, so these black lines represent the paths that the traffic speed cameras in New, York City show all right, so all those black lines.

A

So we pull all that data into new pic and.

B

On just that, little that little section right there that's lit up and thankfully and and I don't want to seem mean, but there seems to have been an anomaly. So we did get to see that and we did get to see how well new pic learns about that.

A

Whenever we run that yeah.

B

A

So yeah it's not going to take that long. So, but we do have a fair we we don't have so the data comes in per minute, so we have a fair number of data points we have just over. We have about 560 data points so when Tom runs this thing we're going to see that- and this is really just for one of those lives. We actually got data for about 200 different routes, but this is just one of them.

A

We kind of cherry picked one that had an anomaly, so we see here going up and down and again this is really just for a very small period of time and just during the weekend. So what we would expect is we would get this data for weeks at a time and then we would see traffic going up and down right now we see where at what two o'clock so then our our cash or kind of crashed, and so that's what this whole line is.

A

It ran out of memory, so we came in the MOT in the next morning and said: oh, that's, that's not right said she'd be coming up. What's it?

A

Okay, so, okay, so there we go so now we got now we go and down at the bottom. You have to take our word, for it is the the prediction: what is it the anomaly so right about there? We got to about 500 data points and then it said hey. We could see anomaly and right here you see this huge jump right here, which is, it might be just around when the weather started. Getting ad.

A

Right- and here we see this flat just until we get about 500 data points, and now we can actually start predicting things. So what we'd like to do, given that the data is still kind of weird, we can continue to run this thing put up a Heroku instance and just start keep keep going going going most of the code. Is there in fact, the code to light up a little line on google maps?

A

Is there mostly, and I think that it would be interesting to see this sort of thing after about a month and see how well it can predict anything? I.

B

Said just as an interesting aside this, this sort of prediction didn't actually take all that much processing power, so we were actually able to run 120 instances side by side before we started slowing down and screwing around with the actual speed of it where we couldn't access it. So, instead of graphing 220 lines, we chose one because it would be a lot easier to detect the anomaly.

A

Right, um that's it! Thank you.

C

Yep, how frequently do you get the data.

B

C

A

The data file is every minute and we're grabbing at every minute. Is it.

D

Changed dramatically every minute, is it a very slow changing over that fridge.

B

um Yeah go ahead. You can see from the from the blue line that it's actually pretty consistent where it and and sometimes like during rush hour, it'll change drastically that it'll it'll go from like we graphed travel time instead of speed, because we wanted something that would go up instead of something that would go down with traffic.

B

A

Well, so so keep in mind that this is just one of the paths um which is fairly flat, the first one we did was more of a gradual up and down. So it really depends on the path yeah.

D

Well, and and again, there's always the risk of oversampling if the day to change his trailer slowly compared to your sampling rate, then- and it's not it's not going to work as well, but you some sort of happy medium. What.

A

We would expect if there was a really bad like crash, that it would go up and up and up and then it would go from maybe 22 boom and those are things we really want to predict.

B

Like like this, that's kind of thing right here, it might actually help to have other types of data sets in the prediction algorithm like taking into account the weather or, if there's a concert or something else, but we didn't quite have the time to implement those.

B

But in the long run, that's the kind of thing that we're going to want to try to implement into the algorithm the.

C

Basic idea of not modeling traffic but the anomalies in traffic flow, that's really cool. It's a nice idea, Thanks.

A

Alright, thanks again.