Name: Structuring and visualizing Data - Deep Learning in Halite AI competition p.2
Uploaded: 2020-03-31T18:20:32.000Z
Duration: 37 min 10 s
Description: Thousands of YouTube videos with English-Chinese subtitles! Now you can learn to understand native speakers, expand your vocabulary, and improve your pronunciation...

Present continuous

Everybody, welcome to the second part of our machine learning with Hallie three tutorial in this video, we're gonna be building on the last one.

Past continuous

So in the last video, we, um we basically were checking t see if we could do this like math of relative positions to get the surrounding coordinates of any given ship.

We definitely confirmed that we could, including getting like the rolling over of the edge of the map and so on.

So we are confident that we can do what we want to do.

So now what we want to do is work on what does the what is the individual ship?

And what is the input layer to our neural network?

In this case, it was specifically a convolution.

So coming on down here to this, like, block of code that we took from the testing grounds that we built, we just kind of needed slightly modify this, but not not too much.

Most of what is here actually gets to remain.

We're gonna be adding quite a bit, though, so for size three is probably too small.

We just did that for testing purposes, I would go with a 15 or maybe even a 16.

If we go with a 15 that would give us a 31 by a 31.

So it's always gonna be this number times two plus one, and that will give you the grid size.

Because if we want to use things like Mobile Net and Inception and some of these other models, the minimum size is 32 by 32.

So we could re size or we could pad or whatever.

Or we could just build data that has the proper size right out of the freaking gate and we'll be all set.

That gives us, at the end of the day a 33 by 33 grid that every ship sees.

So on the smallest map, it will see the entire map on the largest map 30 33 by 33.

For whatever reason, it's probably already lost any ways.

So 16 by 16 For now, surroundings is currently empty and later we were gonna populate with coordinates.

That was just for quick testing purposes.

Now we want to do is populated with meaningful values.

So what I propose is that surroundings contains It's a list of lists where each list is your pixel value, so to speak.

I think my I've lost my That's unfortunate.

My caps lock light on my keyboard no longer works.

I don't know if I could go on in this tutorial.

My number, like all those lights aren't working anymore.

Anyways, eso So the first value will be how light amount, Um, Then ship and then drop off.

So probably will be a shipyard, but it could be, um it could be a just a simple drop off location.

I doubt we will make it that for we will probably only have, uh, shipyards rather than adding more drop offs and stuff.

Okay, so, um so that's what we want, Purcell.

So for each pixel, let's say in this image that we're going to draw the first value.

So So it's we will be able to depict this as if it's our g B.

So we could show this in, like, open CV or something just to test.

But anyway, that should be all the data we need.

The only thing we're kind of missing here is what turn number are we on?

So we have really At this stage, we're not in putting to the neural network.

So if you're like, really early in the game, you could travel pretty good distance.

That's fine, as long as you and the other thing we don't know is like, how much?

How light do we currently possess right now?

So in the early game, or like kind of early mid ish game, it's okay to travel long distances, but towards the end of the game, if you travel out too far and you have time to come back and deposit the end of the game.

So So we have no concept of where we are in the game or how much how late we actually have.

Um, so that's kind of unfortunate that our model won't have that data later on.

Maybe we could also pass turn number or even a ratio of, like, turn number out of how many turns we know there will be.

So you know exactly how many turns there will be by the size of the map.

So you could maybe, rather than three channel continent, you could have a four or five, so we could have turned number and Hal i TTE And then maybe, you know, make the max How light?

Something like that is a pretty high threshold.

So it would still be between zero and one, which you want for scaling purposes.

Anyway, for now, we'll stick with the three channel, But there's every room, toe, ADM.

Okay, So to know, certain things like ship and drop off we want to know is that our ship, or is an enemy ship is it our drop off location or an enemy drop off location.

So before we generate over ships, we do wanna have, um we want to know where are we want to know our ships.

And we want to know, um, we would like to know our drop off locations, so at least our drop drop off.

Um, come, um, I'm trying to decide how I want to do this drop because, like the drop off ourselves, we really could just go me dot Get?

I think it's me dot Get drop offs like drop offs equals me dot get drop offs.

But I don't recall if Mita get drop offs.

And this I don't know if this is a list, so let's make sure that's a list and then plus, um ah, me, that shipyard, we'll just do this.

So this should just be drop off objects in shipyard objects list be dot get drop offs.

Ah, And then for now, let's just say ships equals, Um, no, we definitely need.

So let's do, um, d dot position for D in.

Okay, uh, because we want to know, Because we're gonna be on Lee looking at the cell, so we definitely need to do it this way.

So it's called this drop off positions, and then we'll call this ship opposite positions and then, yeah, we'll just do what we had done before.

I can't remember if this was in our text based version or not.

Subtitles ListPlay Video

Structuring and visualizing Data - Deep Learning in Halite AI competition p.2

stuff

constant

basically

positive