Everybody And welcome to another deep learning in Hal I TTE video in this video, what we're gonna be doing is checking out the kind of results of everything up to this point and then talking about moving forward.

Eventually, we have to stop this project because it's just gonna take way too long for it to have become anything really, truly competitive.

But I still think it's a really fun thing to just see you how far you can get from something like a random.

So, uh, let's talk about it.

So what's coming here first and check out some of our replays.

So this is from these air replays from phase three's model again, All random least to start.

So let me bring in just the most recent replay here.

We got a bunch of replays, but I'll just bring in a couple.

I don't want to.

What has happened?

Um, did I click play?

I probably play.

Okay.

Um, at least purple looks pretty good.

Let's speed up, Purpura.

Speed up the game rather.

All right, 10 x Okay.

Purple li pink player appears to do pretty darn well.

Green player does not look very good.

And, um, light green slash yellowy green player also doesn't look very good.

Okay, so But again, these are all the same model.

So the fact that we have that much of a delta between the players is kind of weird.

Um, the next thing I wouldn't mind checking into is, um, like, if you click on any one ship, you can see how much how light this ship has.

So in this case, this guy has got almost a thousands.

Got 987.

How light he needs to deposit that.

You're just wasting time here.

This ship has zeros.

That's good.

265 Not bad.

853 needs to go deposit 3 30 maybe should deposit.

But who knows?

Let's check these guys.

This guy's only 1976 again.

Needs a deposit for 62.

Quite a bit.

9 86 needs a deposit.

1000.

You know, Green player.

If Green Player would have just went and deposited everything as thegame was coming to a close green player would have won.

Okay, So as you'll probably see in a lot of replays from like top players at the end.

All the ships just rushed the report because it doesn't matter if your ship sinks at the end.

So the quickest way is just kind of suicide on your drop off, and so you we could incorporate this into but into the game at the end or something.

But the real big problem, I think is like, even like halfway through the game, like when you shouldn't be rushing the port necessarily.

We can check 500 right?

Should deposit 972 Should deposit for sure.

891 should deposit 51.

Um, purple actually did pretty good.

So we're not even gonna bother wasting time.

But again here, 8 46 needs a deposit.

These two guys were finally continue along 6 13 Probably should deposit 1000 deposit it, So OK, so I think the thing I would like to do next is probably Force Force ships to do that.

This game actually looks pretty darn good.

Um and also, I know we're ending games not very good like this in terms of the actual competition, we would lose for sure, but in terms of an A I that's trained completely from random toe learn to do these things.

I just think this is the coolest thing ever that has learned Thio slowly pick a pallet and slowly drop it off.

I just think it's cool.

Okay, so, um great.

So let me close and don't forget.

Like at the beginning, random was like collect 100.

How light after 50 turns.

So, in theory, maybe you could extrapolate that Telex a 500 turns and they collect 1000 Allied or something like that.

But that doesn't appear like in this case, we were clearly way beyond that number.

And my guess would be you wouldn't actually randomly collect 1000.

How light after that many turns.

But who knows?

Price should check that out to, um regardless definitely learning stuff something anyways to get an edge.

And some people have asked like how we won last year in the M l bought Well, basically, we had a rule based bought that was really good.

And then we used deep learning on the rule baseball with some random motor movements, just like what we're doing here, and that was able to actually outperform the rule baseball it was trained on and also get number one ml bottle.

And that's all we're doing here is we're just slowly get ej ej ej and so on.

So anyways, but how could we fix at least like some of these little problems that we see?

I mean, there's, like, two major things that we can do here.

So, you know, you want to make the decision.

Do you want to make a rule baseball?

And, um, you know, just have our basically Do you wanna add rules to your bought and then be done with it?

Or what you also could do is you could add things to your body and then those things that you added add them into training data.

So, for example, do exactly what we did here with naive navigate.

So we're trying to train the Aye aye to stop running into itself.

So we use knave safe, right?

But we threw that into the training data.

We said, Hey, by the way, use this as training data.

If we needed toe not move the way that we were planning to move or whatever change the training didn't make that the new input data, so maybe the model could learn to not run into other friendly ships or even enemy ships, for that matter.

So anyways, so we can take that same thing and apply it here with hurry up and go deposit.

So what I'd like to do is, uh, let's just clothes these clothes out of here.

Let's, uh, first I'm gonna copy paste, and then we're gonna rename this toe m l to drop off.

Uh, then before I forget, let me change and run game to drop off here.

So it's a man.

Okay, go away.

S So we're gonna replace this, uh, control H coming out here.

Paste replace all save.

Great.

Good work, team.

Okay, so now what we want to do is I'm gonna get rid of these This comment here.

I don't need that.

First, we're going to clean everything up.

No, just kidding.

Eso will come down here, and basically, I'm trying to get, like, the cleanest way to do this.

But also the quickest way to do this.

Um, let's start by throwing an if statement here, if shipped up Hal, I t amount is greater than or equal to a constant stop.

Max underscore, Hal.

I divided by two.

So if it's anything more than 500.

How light?

Hey, we wantto navigate to that shipyard.

So how would we do this?

Well, um, I vote.

We basically just do this.

Copy.

Pasta done.

Okay.

Not really, but almost, uh, all we need to do is here, rather than, you know, do the position and all that.

What we're gonna naively navigate to is just me dot shipyard dot position Later, If you had more drop offs.

Like if we decide to add them later on toe, like, you know, have rules to convert ships to drop offs.

We could go grab some of one of the drop offs in the drop of positions, for whatever reason, or pick a random one or whatever, but for now, this is what we want to do.

So again, this stuff here is all this This in theory, we can save this choice, right?

We can actually a pen that to the training data right, which were doing right down here.

Now, the other thing we need to do is if if this is the case, do that l If this is the case, do that else do that.

And then otherwise we're doing this silly?

Mmm.

Um, this secondary.

This is not the smoothest way to do this, but just to save us from running naive, navigate twice and slowing down our code.

I'm gonna copy this here, paste.

Okay.

So if random Choice and not in safe knave will do this And also, if we make a model choice, we also want to run it through the safe now.

But this way, we're not doubling up on that safe now.

Also, that would cause confusion also.

Why is this tab ugly?

Are you kidding me?

What is he?

Is this what?

Why?

Why is this happening?

Why does that keep happening in den using spaces?

Stop it.

But what does that do that to me?

Cool.

Okay, we fix that.

But why?

I don't have to keep making that change.

And also, why is that not default?

Why is the default tabs the heck out of here is probably going to get angry at me if I don't fix these.

That's frustrating.

Okay, good.

The ro set.

All right, So now when the ship has more than 500 Hal i TTE, it will go and naively navigate its way overly to the shipyard.

Let's see.

Uh, Okay.

So what I'd like to do is we're gonna take replays.

Make that replays old.

Let's make a new replays re plays.

Then we're gonna run game, Python run game.

Hopefully, no errors.

Come out.

Okay.

Making it through some turns.

All right.

Okay.

Looks like we're good.

So what happens?

Um uh, replays end.

I've already closed the watch, so let's go ahead and open up, Hal.

I dot io watch.

And I might I might have to, like, pause or something, actually come back when we've got, like, a healthy number replays because not every replay is gonna be very good.

And this a eyes pretty bad when, like, there isn't a good amount of Hal I TTE in the vicinity.

So, um, so it's kind of by chance if it's a good one or not.

But anyways, let's see how we did.

Okay, so this one's not bad.

Do these oranges looks really good.

I don't know how many ships orange ended up producing.

We really like high ships, but we've got some seven ships over there.

It looks like orange maxed out at five.

But, um and there's a lot of GM is happening here.

I don't like that either.

My guess is I wonder what those guys have in terms of how light.

Well, we'll check them here in a second like this poor soul locked in the middle.

He's at zero, though.

No big deal.

Any of these guys over 500?

No.

Any of these guys over 500 eight?

No.

Okay, cool.

How about green guy?

Don't worry.

While we're checking this, Bt dubs, uh, we got other games running in the background.

2 47 Okay, great.

Uh, okay, let's see what this next game.

Did you.

This one looks even worse.

Rose Home Mart ships in this, and this should have been a, uh, sort of in a pretty good game.

We've got a lot of lot of good.

How light?

Right here.

I wonder.

Like which we could stop the, uh I wish you could stop the back and forth.

I wonder, too.

Like, I wonder if you know the fun.

Uh, if that What if We also said if ship dot position equals, need a shipyard dot position, um, copy paste.

Because we have a lot of these ships that keep finding their way on to So I didn't want her.

Like what if we kick that ship off?

Uh, if go.

And so what if we make this?

I guess we could make that an LF things getting sloppy.

But I'm really curious to see if we get, like, constantly boot ships off of the, uh, the shipyard.

Any drop off location?

Really?

Uh, so you can continue doing stuff like this.

You also could, Like I was saying before you could add in some rule based things, like when you know how to get head towards large patches of Hal Aitor.

You know, something like that, and then continue coaxing.

You're a I tow, like, learn the things that you want it toe learn.

Also, our neural network is pretty darn basic.

We're not really doing it a fancy a eye at all.

Let's see how this one did.

Hell, I might even be worse.

At least they stay the heck off of it.