Numenta 2014 Spring NuPIC Hackathon Demos, 6 May 2014

Previous Meeting

⏯

youtube image

►

From YouTube: NuPUCK: 2014 Spring NuPIC Hackathon

Description

Hack demo by Steve Levis. Source code: https://github.com/levis501/nupuck

A

Great all right, hi.

B

A

Thanks for uh coming, I.

B

A

What I thought I'd like new pic to do is um predict uh the future state of a game and make changes to sort of a primitive program of behavior based on those predictions.

A

What I uh the game, is a shuffle puck or like a two-player pong that you may be familiar with. Let me just I'll just start the demo.

A

The the first thing I did was I wanted to generate some data, and I wanted very, very simple data. Like I said, I'm not too familiar with new page just still. Learning um the data is a sort of a one-dimensional field of view of the vision of uh the puc, uh the the the player on the bottom of the screen. The player at the top is the ai that came with the uh that came with the game um and below the score.

A

You can see these two little two yellow bars, kind of moving back and forth, and that's the field of field of vision from the point of view of the the bat on the bottom and that that, where those bars are, is this data that you see streaming by here on the left um and that's the data I fed into uh swarm?

A

uh Actually, I I ran two swarms, one for the to predict the the left eye position and one to predict the right eye position now right now, uh this okay, I have a diagram that will uh explain that a little bit better.

A

I guess I need to exit here.

A

um This is a schematic of the game board you have. uh This is where the top player can move this but bat around here's, the bottom player- and here we have a a puck- um and uh I imagine, uh there's an eye on either side of this bat down here. Looking at the puck.

A

This is this is a. This is one of the eyes, and here the the puck is a little to the left of the eye and it's kind of close up to the to close up. So it's you know, it's uh you know takes up three of those little squares and it's off to the left.

A

um If it was off to the right and say a little farther away, um it would be smaller and it would be off to the right and then I uh that what you're seeing in those yellow what you're seeing in those yellow bars is both the eyes superimposed on top of each other.

A

um I originally imagined just converting that to a series of zeros and ones and using that as input to spatial pooler, but uh I I ended up using floating point values with the swarming, uh because I was able to get that to work first and then so I I ran like I said.

A

I ran two swarms, one to predict uh the let where the left eye position is and one to predict where the right eye position, where the puck's position is within the field of view, is what I mean to say there um and uh and what uh what I I'm predicting um a hundred time steps 100 frames of the game in advance, and I can't right now. I can't tell if it actually does much better, um although he actually just won a point there kind of happy about that.

A

um uh What uh probably I would in the future, what I'd like to do, you know maybe, is figure out how to better incorporate the prediction into the movement of the of the bat, um the bat the eyes I are moving with the bat.

A

So if you pre, if I'm predicting, that the puck is going to be right in front of me right in front of the bat, there's no reason for the bat to move. However, the bat is uh has a sort of a primitive programming where it it moves kind of along exactly what it sees at the current time. So if it sees it's the left right now it moves.

A

But um if it's predicting that it's going to be right ahead of me and I'm using that prediction to move the bat, then it's not going to move at all.

A

I'm thinking, maybe in terms of incorporating the prediction into the movement of the bat, tries to try I'd like to try a bunch of different parameters like uh some different thresholds say if it's really far off to the side, I want them. I want to move- or maybe I otherwise wouldn't have moved. um Oh I've scored three times already all right. I'm pleasantly surprised.

A

uh Yeah I uh I'm running um in python and a separate process that you don't see here. um I got both the models fired up and it's receiving uh the data over uh over a socket and and sending the sending the results of the two models back and then, like I said then incorporating it to the movement. All the behaviors.

B

All the behaviors generated from new pick at this point in time. It's not a combination of the old algorithm and.

A

This is I I well I when I first did it it's a combination. It's um it uh it. Basically, new pic will override the default behavior. If the. If the prediction is saying it's going to be farther out than it's seeing right now. So if right now, if the bat sees you know sees it, it's uh you know pretty near the middle and the prediction is saying it's going to be far off to the side. It actually uses the new big prediction to move.

B

So you can't really tell from this. I can't tell if necessarily how much nupic is contributing versus how much the old object.

A

Well, uh no, I think I would need to run many of these uh uh many runs of this here to see if it was doing better. Although I have to say this is the first, uh with the without.

B

A

New pick, it would score maybe once or twice and the other one would score all 15 points and win the game.

A

Yeah uh sure that would be a good like good statistic to keep an eye.

B

On the question.

A

Does it move only two dimensions because the aaa actually moved three dimensions right? Oh.

B

Yes, I I limited.

A

Two dimensions: just to uh simplify debugging yeah.

B

A

Actually, um like average, because it moves really fast, do you average the uh I mean the predictions arrive really fast. So do you actually do an average during time, so it doesn't do those weird movements, this that kind of jiggling back and forth? um No, I I I don't um it's just it's a it's an all or nothing thing and again that was mainly. I can't believe it's still winning. Okay, uh it's it's uh uh uh yeah! Yeah! It's running! We can see.

A

It's uh I'd say: if I can stop this a little bit.

A

Oh never mind this, isn't the right data.

A

But yeah I.

B

A

It it is running the prediction: it's doing pretty well yeah, I'm uh quite happy.

B

A

I I there's some improvements. I think I can make in terms of uh making sure that uh you know it's it's doing as well as it could and um just different ways of using the predictions to augment the behavior very cool all right. Thank you.