Placeholder Image

字幕表 動画を再生する

  • What's going on?

  • Everybody welcome to another part to the M l in hell like three.

  • Siri's in this video, What we're gonna be doing is checking out our distribution for randomly Trick or a model that learned from random movements.

  • But of course we set a threshold.

  • And the question is, Did the distribution of all games above?

  • I think it was either 4100 and 4200.

  • Is that distribution after the model's been trained?

  • Has it improved?

  • We're not expecting to see a huge improvement, but some improvement would be nice.

  • So Ah, what we can do is open up training, data distribution.

  • This reads training data, which is the one that we made.

  • Let me just take this copy pay.

  • So the old one, our original one averaged a 42 02 and the new one averages a 42 96.

  • So quite a bit better.

  • Almost 100 about 94 better in, like 100 out of 4000 would be 2.5%.

  • So maybe somewhere around two ish percent is what we got here.

  • Improvement over random, which sounds horrendous.

  • But if we can continue to get that kind of incremental improvement, then we would There's there would be something there for us, and, um so I think, you know, this is a worthy practice to continue doing.

  • But I'd like to make a few changes to our our actual bought here.

  • So, um, there's a few things.

  • One is we're only playing for 50 turns.

  • Still, I still want that to be the case, so long as our A I is relatively dumb, there's no benefit.

  • I don't think through running through longer games, So I think we still want to keep at about 50 turns.

  • I think that's a pretty good number.

  • Maybe we could raise it to 100 or something like that.

  • But I also think I want to start this model more again from scratch.

  • So, um, because I just wanted to validate my question.

  • So now what we did was we just went based on total highlight collected.

  • But the problem is, every game map is random, and it's more likely going to be the case that the, uh, a eyes that did, you know exceptionally better just did better because there was more highlights on the map and in general some of the maps just have less Hal, I TTE some have more.

  • So what we need to do is come up with a way to handle for that.

  • I think the best way is to gather the initial amount of how light and then compared to the end of the wherever we are in the game, how much Hal ites plus each ship.

  • You know, what is the percentage of Hallie this a I has at the end of 50 turns, I think that makes a little more sense.

  • So that's what we're gonna d'oh.

  • Uh, the way I'm going to do that is, uh, first we need to gather all of the hal I TTE in the map.

  • And, um, it's been a while.

  • I don't think we actually have straight how light either As a highlight map.

  • So what I'm gonna do is I'm gonna copy and paste a bit of code.

  • Uh, if I forget to put this, at least I'll put this in the description.

  • I'm getting way behind on text based versions of editorials.

  • I apologize.

  • So it is.

  • But I can definitely put this snippet in the description, so if it's not there and we're watching this or whatever.

  • Especially if you're here for the premiere.

  • Let me know, because I forgot it.

  • Hopefully I'm here for the premiere, and I checked right now to see if I have it.

  • Anyway, Uh, what we're gonna do is I'm gonna copy this code.

  • I'm gonna come into the while.

  • True loop here.

  • We probably just need game map is my guest.

  • Let's see.

  • Game map yet?

  • Yeah.

  • So do paste that in eso.

  • This is code from Daniel.

  • How did write this?

  • I would definitely write a much better way.

  • I would never use this whole like, if none Because we're iterating over a loop.

  • I mean, horrible, horrible code.

  • I'm just kidding.

  • Okay, so, uh, this is our way to figure out what?

  • What did them What?

  • What?

  • How much how it started on the map.

  • There we go.

  • Got it out, Uh, as my phone again, Not yours.

  • Okay, So, um, So we know how much Hal I started on the map.

  • And then the question is, what percentage of that are we collecting?

  • And this should be a little bit better of a way to do it.

  • So now what?

  • We're gonna do is basically the end of 50 turns or question to save, uh, due to do to do great.

  • But here is what we're probably looking at.

  • Yeah, um, our question to save shouldn't like save threshold probably isn't the right answer.

  • Instead, what I would do is something like me dot Hallet amount.

  • Um, plus, uh, Len me dot Get ships me dot Get ships times.

  • Let me times.

  • Um, constance constance dot ship cost.

  • This is gonna get kind of long.

  • Apologize.

  • I'm gonna put this imprint seas.

  • I believe Pem does would save us there, and we'd be okay.

  • I'm gonna do that anyways.

  • So how much?

  • How late we have right now and then we're gonna say each ship is worth 1000.

  • How light collected.

  • So that number.

  • And then I put all of this in parentheses.

  • If all of that out of initial hal I TTE, uh, due to do to do is get back up to this initial how?

  • Let him out here, divided by initial how light amount.

  • Let me just full screen.

  • That's no reason not to, um is greater than or equal to save threshold.

  • So this is gonna be some sort of threshold I don't really know what to go with for.

  • So I'm gonna go with something really small because, actually, I just don't know.

  • So save threshold.

  • Um, let's just make it really, really small like 0.1 Okay, eventually, hopefully, we can get more of like a, you know, 10% or, you know, something like that would be really nice, so we can hope you.

  • Okay, Now, um, let's make that say that to a variable so we can save that.

  • Our file, actually.

  • So let's say, uh, percent gathered equals this, huh?

  • Oh, no.

  • Did I forget?

  • I know.

  • I guess I didn't.

  • So it's really all this boop.

  • And then do this percent gathered pace that there in that correspondence school?

  • Uh, if percent gathered great.

  • And then rather than saving me dot How light amount, Let's save percent gathered.

  • I think that's a cooler, cooler option.

  • Also, just for the sake of it, let's say, uh, ships equals and then I'm gonna take f uh, let me die.

  • It's a shoot.

  • Okay, this is for the lens.

  • So we're gonna say ships, so we'll call this ship.

  • Count on, then ship count.

  • Okay?

  • And then because I also want to save percent gathered, um, let's save ship count as well, because later we might want to use some sort of logic that balances those like we might wantto incentivized more ships at the end s O, for example.

  • Right now, we still don't have safe navigation.

  • So these ships air highly likely to just run into each other like idiots.

  • So maybe if we see that's a problem, we want a rather than worrying too much.

  • About percent gathered, like maybe for a next generation.

  • We actually want to keep percent gathered the same.

  • We're not going to try to raise that bar, but it's still we're gonna do is try toe incentivize a eyes that have more than two ships remaining or one or whatever.

  • Okay, so, um, percent gather on.

  • We can always do that after the fact, BT dubs.

  • Right.

  • So after the fact, we can always parse these files split by dash and go with whichever one we want.

  • So So we don't really have to do it here.

  • Eventually, we could also make that some sort of threshold like must have more than three ships and must have some certain percent gathered threshold.

  • All right, now, before I screw up, I just wanna Yes.

  • So this is our old data.

  • So I'm gonna make this training data underscore to underscore, too.

  • And then we'll make a new one called training data.

  • Underscore three now, hopes not.

  • Eventually, we'll call it that.

  • But right now, it needs to be just training data.

  • Cool.

  • Okay, next thing is, we want to change max ships.

  • So, um, where do we set?

  • Make sure.

  • OK, one idiot.

  • I'm scared.

  • Said that's, like 99.

  • Eventually.

  • I just don't think we should have any number of maximum ships.

  • The thing that will want to handle for later is in turns.

  • So usually, after about half of the turns have gone by you, I want to stop producing ships and just simply collect allied eso eventually will add that in.

  • But right now, as long as we're truncating the number of turns, it doesn't matter.

  • So, uh, in fact, it might even be the case that we should never do it.

  • I really don't know, Um, and then on Lee, when we actually submit the bought for the competition.

  • Do we stop that process?

  • But would it really hurt toe?

  • Have the model be trained under that man?

  • I'm just not sure it would make a difference off anything.

  • Maybe it makes it better at, you know, avoiding its own ships.

  • Um, okay.

  • So dare I say we're done?

  • So this might be one of the shorter ones that maybe I'll clip, you know, appended this to another bit.

  • We'll see.

  • So, uh, we'll see if it runs.

  • I mean, uh, TV Python run game.

  • Let's go.

  • If we hit some errors, it's taken a while.

  • Oh, Well, OK, there's a couple of things.

  • One is We're probably loading the model.

  • I'm not really sure I want to do that.

  • I don't think it really matters.

  • We could load that model.

  • That's not a big deal.

  • Yes.

  • So we're gonna load that Phase two model.

  • I don't think it's running Now.

  • Let me think about this.

  • What else?

  • I'm trying to think of what we hit along the way.

  • So turns yours or one eyes.

  • Could cirque.

  • Uh, that's valid.

  • I think we can all agree.

  • Unless unless But the vat, the variable hasn't been defined so it should ring as a nun.

  • But maybe that's maybe that's been the problem all along.

  • What?

  • Nube.

  • Okay, so I guess I just need thio.

  • Set it to none.

  • Initially, for some reason, I was under the impression that you're completely allowed to, ah, reference a variable that doesn't exist.

  • And if you do that, it's called None.

  • But clearly, I'm not getting away with that, curiously enough.

  • Uh, okay, so the more you know.

  • All right.

  • Uh, you do.

  • Okay.

  • So now what we want to do is come into training data, and then let's just delete all this.

  • The first thing was percentage collected.

  • Writer is the first thing time that we're saving.

  • Okay, Percent gathered ship count, blah, blah, blah, blah, blah.

  • Okay, so these are our percentage gathered.

  • We are always ending with one ship.

  • No surprise.

  • I am sure we are just crashing into ourselves constantly.

  • That would be my guess.

  • We can find out.

  • Go into replays.

  • Surgery here is, uh, just it's getting kind of Laghi up in here.

  • Play.

  • Yeah.

  • See, we're just producing five ships very quickly and crashing into ourselves.

  • Look at the abundance of Hal it on this map.

  • Who?

  • This should be a juicy, juicy map.

  • Unfortunately, it's not.

  • Uh, So, um, I think you know, random will save us, and eventually we will will end with two ships.

  • My only fear is that, you know, whoever happens to collect back to it, although even collecting back to 1000 would be fine.

  • If if they do that, I think we're good with this model.

  • Over time, I'm just really hoping eventually we will end with, uh, enough samples of two ships.

  • I mean, so far, we still don't have any and nowhere near even, like, 1% collected.

  • That's kind of embarrassing.

  • Anyway, Uh, okay.

  • So probably what I'm gonna do is let this go for a little bit, and I'm gonna see if we get any samples where some ships are retained.

  • Um, because we really we were you know, we're not really actually setting a limit on anything else yet, So I'm thinking, uh, maybe I'll end up with some sort of, you know, yeah, percent collected threshold and ships remaining fresh hold or something like that.

  • But before I can come up with a good number, I need to get a good sample of data just to see Do we ever randomly retain any ships?

  • And if not, then we really need to probably safely navigate.

  • Unfortunately, I don't want to do that.

  • I don't want to force that on the model.

  • The good news is, if we do naive navigate, it does return the move.

  • That is correct.

  • So then we could still use that to train the model.

  • So we could so feed that in as training data to the model and hope Hopefully it could learn to the point where we could remove the naive navigate, but it's not looking good, fellas.

  • Anyway.

  • Okay, I'm gonna let this run.

  • I'm gonna have a couple of our view of him running at the same time on we're gonna see if we can't get any games where we get any ships retained.

  • Uh, what a bummer.

  • Surely in, like 50,000 samples, though, we'll get some will see while we wait.

  • Uh, let me give a quick shout out to my recent channel members.

  • We've got Jeffrey McGill, ski rish off a Nonda, Ravi Krishna and Michael's.

  • Severson.

  • Thank you guys very much for your support.

What's going on?

字幕と単語

ワンタップで英和辞典検索 単語をクリックすると、意味が表示されます

A2 初級

モデルの改善点を分析-ハライトAIコンテストでディープラーニングを実施 p.7 (Analyzing model improvements - Deep Learning in Halite AI competition p.7)

  • 2 0
    林宜悉 に公開 2021 年 01 月 14 日
動画の中の単語