Everybody.

And welcome to part seven of the highlight three tutorial series, as well as the first part to the M l section of Hal I three tutorials.

I do want to stress I am here mostly just to tinker with M l.

I have no promises here.

It's kind of like the python plays G t a or the S E to Siri's.

I don't know what the outcome's gonna be.

It's just a realistic example of me trying to solve a problem with ml eso.

In one case that was pretty darn successful.

And the other one, it wasn't so.

So we'll see.

If you guys have ideas, how to make it better or whatever, feel free Thio.

Leave it in the comment section.

Also, if anybody has an ML type of model that they just don't have the processing power to go through if you want, you can submit those in the comment sections, too.

If I see something that looks pretty darn cool, were interesting all happily devote some processing power to it toe, see if it's ah any good.

So if you want to do that, have at it so with that.

Let's just jump in and let me talk about what I want to do and why.

So the first thing I'd like to do is, um let's head over to highlight dot io and talk about the method I want to use.

So one method that is probably gonna win Hal I TTE or at least be the best type is you can parse replay files now.

I'm not gonna totally hate on that method.

I plan to use that method.

Um, and I think that will be the one that wins if there is a winning ML method.

But I find that method not like it's cheating, but it's kind of cheap.

It's not as cool to me as something a little more raw, so that's a very supervised ML method, and I think it's way more interesting to do methods that are a little less supervised.

So what I would like to do is the type of model that I always want to D'oh, which is a more evolutionary based method, something that learns more from like Random is just fascinating to me.

I just really like that.

So that's what I want to do here and So anyways, I know people are gonna want, like reinforcement learning.

This is to extent kind of reinforcement learning.

But it's not Q learning, which is really what everyone means when they say reinforcement learnings.

Anyway, uh, we will be doing deep learning.

Just I want to do an evolutionary based model because I think that's really cool.

If you want to do a reinforcement learning with Q learning example and again, you don't have processing power, let me know.

Post it below, and I'll happily see how good it is.

And I'll dedicate ah, comparable amount of processing power to it to pay how decent is so on.

Give you credit and all that.

I won't submit the Bader anything unless you want me to.

Anyways, let's watch a game.

Um, I can't click it.

They're going so fast I can't click thistles.

Super frustrating.

Thank you.

Yeah, I couldn't click on him.

Okay, so, uh, here's an example of a game.

And here's how I how I want our approach.

It in his few words less.

It's gonna be a lot of talking.

I'm sorry.

There's a lot of concept here.

Um, so anyways, so what's going on?

here.

Okay, so in this case, each player has what?

For ships?

Four or five ships.

Think green only has four.

Yeah.

Anyway, um, each ship is its own little entity.

And in the world of Hal, I TTE this game, as I've said before, is really all about navigation.

But the nice thing is, in highlight to navigation was you had a vector and thrust and all that's like a percentage those very challenging ML operation was and how it three, Honestly, this is a much more basic problem to solve.

That doesn't mean that competition is any less that we've got a lot of the same players from last year and the competition is Justus.

Hi, but the actual challenge itself and the way you need to think that part's a lot more basic you Instead.

Now you've got to get a lot more fancy when it comes to actual logic and stuff like that.

So I'm definitely dogging on this year's game, but in terms of navigation, we only have five options, right?

You can only move up, down, left, right or not move.

So that's pretty nice.

And it's a grid structure, so it's very basic It's super simple for us to feed this information to an A I or an M l A I and specifically a convolution Eleanor Network.

It seems like a no frickin brainer to feed it through a confident.

So the only fear I have is they're pretty small, but, uh, you know, we don't need a huge confident we can use a smaller, confident and be just fine.

Okay, so I want to see it through a confident.

Now there's a couple of issues we have.

One is the game sizes tend to vary a few different ways.

We could overcome that.

We could either use padding or we could crop down to the smallest size, which is a 32 by 32.

But if we're gonna do that, Theo, other option is to instead of taking the entire game map, we can take things relative to a ship.

Now, we're already gonna need to do that.

So consider, you know, I'm gonna zoom in, but consider you know, this is your game map.

Let's say it was a smaller game.

APS This one's already kind of big at 48 5 48 which where should this ship go.

This ship could should clearly go left writer or west to get that.

How light spot.

What about, um what about this ship?

Should this ship also head to that position?

Probably not.

This ship should probably maybe even come down here.

It's over here is a little close to the enemy because it rolls over.

Well, it's actually not close to you and me.

I'm zoomed, and I forgot about that.

So maybe he should go over here because he he is closer than the enemy is to this spot.

Right?

But regardless, each ship is relative, so each ship should have a relative viewpoint as well.

So whether you wanted to use the full map or a slice of the map or whatever, you want to not have this, let's say you're this ship.

You actually want the map to look like this.

Unless they were at this ship.

We want the map to look more like this, right with your ship.

The one you're considering at the present moment in the center.

Unless you were going for, like, a macro model of some kind to figure out.

Okay.

Were we send clusters of ships or something?

like that, which is totally valid.

But what I want to do is an evolutionary based model, and I want to do it on a per ship basis.

So each ship, the model will run a prediction for each ship as we generate through.

Maybe later on, I'll change that.

But to start, that's kind of what I want to do.

So then the next thing I want to dio is we need to get all of the information around that ship.

So we need to get this Coordinate this, coordinate this one and this one and just do that around that entire ship.

So now let's talk about how I plan to do that.

Because the good thing is, we have relative and game map coordinates built into that little Halle A p I for us.

So doing this is really just a logical question.

It's not gonna require too much digging on our on our on our end, so I'll just minimize this will come into here.

Um, and for now, I'm gonna copy paste center.

But I'm gonna call this testing grounds.

Open that up.

Let's just talk about how we can get those coordinates.

So, first of all, we just kind of want to get, like, a radius.

I'm not gonna get a circle, though.

I'm gonna get a square.

So radius is probably the wrong word to use someone say size.

Let's say the size we want is three.

So three in all directions.

We like that square.

How would we do that?

Well, what we all we really need to do is, like, four eye in there.

Actually, that's a four X in range of negative one times scoops, not three size.

Let's make a dynamic here.

Um, and then you need to go up to size, but the weight range works is it starts from a point and goes up to not too.

So we won't say size plus one.

So if I print X here, we'll see you.

We go from negative 32 Positive three.

So if we have a three by three cents because of zero right center position, it's not gonna be a three by three.

And because it goes from negative positive, it's also notice six by six, this will generate a seven by seven.

Okay, So, um, so that we could just do the same thing for Why?

So let me just copy Come down here paste And then for why?

And then we just print Let's for now we'll print X Y and this should be our relative coordinates.

At least it's every combination of relative coordinates, But it is that what we want?

Well, looking at this, I I would say no because this was this was this is like if we were to pictorially represent this like so if we were to get the hallowed information and then pictorially represent it, it wouldn't look like this because of the way we're iterating.

And I think just in case someone's not understanding it, well, we definitely what we want.

This is gonna be the most beautiful stuff you have ever seen.

I'm gonna draw a bunch of these.

You'll have to excuse the mark.

I'm not gonna waste too much time, but this is something it's always like Whenever I have stuff like this, I always have to look at it visually.

So here, let's say our ship, Let's just d'oh!

Let's airship is we'll go with green and we're right here.

This is our ship.

So ships at center, then, like we want the square.

So since it's a three, um, in theory, I guess we want, like, this thistle.

We want this square, though.

But what are the coordinates at these other locations?

Well, so this is up one.

And I want to say up like up is actually negative, right?

And then left would be negative as well.

So at this point here, this is the same ex.

Seems to be a little bigger.

That was my phone, not your phone.

Um, this would be same XO at zero.

But then the why would be a negative one, And then this would be again same zero, but negative too.

And then this would be, um zero.

Negative three.

Now, if we continue, this would still be a negative three.

Why?

But this would be a negative one.

Negative three.

And this would be a negative too.

Negative.