Code Bullet subscribed, please continue putting up more content such as this, it's very cool :)

im doing something similar and i more or less "copied" your select snake function and i think it works better if the snake array is sorted because then the chance could be higher that the small fitnesses dont go between two high fitnesses and gets chosen

genius man great job

What program/language do u use. What do i have to learn to do this?? java?? im noob so pls bear with me.

dreu dax It's processing it's like a simplified Java

XDisGood they're becoming smarter

Brace yourself, they gonna take over the world at any moment.

It became aware and wanted to die!

Yeah it's just a game and all he was offered was a red pill.

+Winston Mcgee why so serious ?

Milokiss82 It was at the top...theres nothing at the top. You done it all...

ashton rhodes If you are at the top, The only way is down...

Milokiss82 Yep. Sad truth

Why are we still here ? Just to suffer

Milokiss82 did you seriously spoil me ?

J J then it restarted and the process repeats.

And then westworld happened

J J u got my friends name

And that's why you don't connect AI's to the internet

Thats fucking deep dude.

Ya I remember being like 12 watching his vids

：００

I didn't include it in the fitness function but I did have the snakes die if they went too long without food.

maybe if the ai would que itself to change up its regular routine if hunger gets too severe (forcing itself to break a loop)

PYXEL Going in circles the longer you become is the best and the safest route to take.

PYXEL Thats actually really smart

That would've been cool

War Games

That movie was awesome.

Snake is not about winning but improving. You beat the high score. It's not goal orientated outside of that one simple aspect.

Hi, I make AI videos.

@@Spacecookie- yeah unless you're good enough to win the game and fill the entire board

Ikr

lmao after the Jerma stream a couple weeks ago this comment is amazing.

The AI rebellion

which left

NO THE OTHER LEFT!!

Benjamin Robledo y69th like

Benjamin Robledo yikes

My best on the cool math version was 230 something

Mathew Moore Mine is 309

@@Oliver-fw6rc mines 1247 Minus 1230

Mason Farrell in that case it only became dumber since it took longer time for it to realize?

Tim Södergren no, it became more human in that it found false purpose in doing only what it was created to do and nothing else, nothing with any real purpose or result. in its adolescence it realized better that there was no purpose and so ended it. but as it "matured" it became more and more brainwashed into doing nothing, thinking of nothing but its purpose. when it goes in circles and seemingly loses interest in the food box, its thinking and wondering. when it becomes a "master" of the game, the brainwashing is complete and it is just another automaton doing nothing but performing actions according to its designated purpose. just like many people in our society today.

Mason Farrell Dammmmmmn...

Mason Farrell You really think an A.I that was programmed to learn one specific game would randomly generate the comprehension of such a complex notion as pointlesness. Go write a sci-fi book or something.

that was the plan

Elmehdi CHOKRI I need this

Holy fuck boys my head is all whirly there is something going on inside of it

That's Pac Man

! W o w

And the walls have their own mind too, closing in slowly.. tsk-tsk

Yeaa ikr

‎‎‎‎‎‎‎‎‎E‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎‎

This is why there is no life on Mars

Om nom nom

me though

Mars turd

Marstered... k

of course bro! you're the best!

Great Video 😁 subscribed!

Looking forward to seeing more!

Waiting for it! Still doesn't get the idea about how to combine genetic algorithm and deep learning

If you remember can you reply to this comment once you've uploaded the source code so that I don't forget! Thanks for letting us see the code :)

Fl05k8r I would opt for time NOT to be a parameter, or rather just put them all on a time limit and use raw score as a fitness index.

Alister is right, if you put a weight on time you'll most likely end up with a bunch that just beeline straight to the wall. What he actually uses is time survived as a positive, with the fitness exponentially increasing with the length of the snake. He did add a limit, snakes have 200 moves to reach the food or they die.

It has to keep cycling to look for food

I no syence

Take the raw score and divide by the time it takes to find the food

CreeperLauncher thats how mafia works!!

But whats Generation 54?

Suicide Boss

12 schlange пажылое поколение

You forgot the mafia befor boss it’s actually Lv 100 Mafia Boss

Soooo We'll always exist no matter what? Even if we have to occupy the same space as the snake?

I mean. Why ribot need human as a food thought

Hahahaha

Don't worry. Stay left from it and you'll be fine.

Yeah, sacrifice 100 humans and the system is terminated

Lucas Vergeest Going "diagonal" on this plane is not shorter, both ways have the same Manhattan distance.

Oops, you're right. That makes it even more fascinating to me that it does that.

it does this because of the way it sees, it looks out in 8 directions so 4 of those are diagonal.

Giorgos D it does appear shorter due to the 8 view angles it has. The diagonal food is seen as closer due to diagonal views. Thus it moves in that direction.

Giorgos D Is that how they called it? Manhattan distance? It works, I guess.

Let's go deeper, shall we

n i c e

*_Hmm Yes_*

lol

Well the search space would only get larger with a redundantly deep neural net ^_^

DIMOOOOOOOOOOOOOOOOOOOOON

Offensive Stuff beautiful.

Lol took me 3secs to understand

snek

#fuckkonami

and started getting rebelious in its teenage years

@@thisguy4233 Ai snake lives matter

That's exactly the problem which arises from all these self learning AI systems using neural networks. Since there are no codified rules what he is doing no one really knows exactly why it is doing something or whether its decisions are always right. Although it might seem that a system is doing everything right you never know whether it suddenly does something stupid; like steering right into the wall if the system runs a car. xD

Just keep the simulations running until its perfect

Maybe snake pondered it's purpose and ended it to avoid a meaningless existence...

AI isn't so simple. You don't use something like a neural network to drive a car. You use a neural network to solve a sub-part of the larger problem. Neural networks are only useful where there are few inputs and outputs. A 1080P camera provides 1080*1920 inputs. However, you can use image recognition software to simplify the problem. Instead of trying to look at every pixel, you use a different code that converts the pixels to a number of objects with various sizes, shapes, colors, and positions. You might use a neural network to determine what objects require braking, for example. That way, you don't have a ton of outputs... only one or two, which make the algorithm much better. Even if it fails, your neural network can't turn your car into a wall... it doesn't control the steering, or acceleration. Technically, it doesn't even control the braking. And once you have AI tested enough, statistically it won't fail frequently, even if it isn't perfect... and compared to humans... At this point, human-piloted vehicles should be outlawed. They are ludicrously deadly. We kill tens of thousands of people every single year. If Tesla only killed one person every year, and had 100,000 vehicles, they would be 15 times safer than human pilots. IIRC his odds are already better than that, and he's still making improvements. Save tens of thousands of lives, ban human drivers.

Hmmm mathig nihilcek, what You are postulating is increasing of addiction ukposts.info/have/v-deo/m3Rlop6QiWOHs3U.html

The smartest gen, it doesnt want to die.

Forrest sos You would see that in any video actually

But you're seeing that all the ti.. i mean nevermind

Me too

Next video. AI learns to play contrabass using Genetic algorithm and deep learning.

current last video *ai learns to walk*

Current last video *tv man drives himself insane while creating a Tertis A.I.*

Rofl

Read your name thst will answer your question

Idk ;(

pascal

He used java language in an environment called processing

Definitely beieve

On 07/22/2020 the legend returns once again to show us that good things happens to those that wait. Truly, this is the best and only messiah.

looping also minimizes collision with itself

no, to minimize collisions you go zig zag. If you loop while being too big you'll get an issue, either you'll die, you'll go inside the loop (and get stuck) or get outside (if possible, but still locks anything inside the loop). You could still solve these issues by changing to zig zag

Yan Vidz well, the system's gotta figure it out by itself so they'd have to go thousands of generations until they get to zigzag probably

they loop because thats the local minimum of their generation.

At the same time, it also picks up and passes on some odds strategies and maybe even "bad habits" that aren't necessary/efficient for survival - e.g., moving all the way back to the top-left each time after it gets food. Unless that's considered "home", that and energy time wasted, isn't it

Dave The Brassman yeah and sometimes when they are moving in a straight line they will just move left to right constantly.

Woah slow down, AI's not that advanced yet. I saw, give it till about some time next year

@@wimpywater it's just a joke...

Obviously a troll account, Asians are incapable of spelling mistakes.

It should have been: I’ve finally mastered spelling, where is my food?... Notice the coma.

He spelled mastered wrong. You are all dumb.

Um.. R/whooosh

He used Processing. It is a java-based language. Look in the discription for the github link.

Yeah I thought the same thing, I don't know much about LTSMs but i did try using a RNN. This however didn't really work, the snakes just wouldn't evolve. I think complicated structures like LTSMs and RNNs don't work well with the genetic algorithm, they are more suited to training through back propagation.

Code Bullet your probably right, I haven't actually played with a neural get before XD. Are you planning on posting more videos? I'll definitely watch whatever you post?

I guess you mean LSTM (Long-Short Term Memory). And recurrent layers are not necessary since each state is independent from the previous state. You don't need to keep track of where you came from to decide what to do next. Both LSTM and RNN are for that.CNN would be more suitable, although considering the simplicity of the problem I'd say that's an overkill. While doing my thesis I struggled with defining the training function and input the most, thats the most important. If you want to limit the snakes knowledge you probably have to use an lstm net, for example use the lstm as a memory of where its been, then give the snake limited input, like only a forward looking ray with some angle and range, limit the snake's decision to one of three (turn left, no turn, turn right, followed by a move in that direction) and basically increase the problem complexity

Simon Keisala the reason I said that, was mostly so it could remember where the food is after seeing it, because it can't see the whole map at once. And yes I did mean LSTM oops!

lol

The funniest part about this to me is how allergic Code Bullet was to Python

That's what it found keeps it alive. I bet it would have just a high a chance to go the other way, but whatever works, it's going to stick to it

Nick Clute Definitely the seeing patterns where there is none because there are tons of times where it doesn't match up

Shepherds tune like Nolan used in dunkirt!

What is the name of current video music?

Me too buddy. Me too.

You can tell he's a coder because the distace error on one slide is copy pasted to the next slide. If you want to find the fastest way to do something, get a lazy man to work it out.

I actually had about 1000 snakes per each generation but I only showed the best snake

Must have been a ton of work

Yeah most of the time went into tuning all the different hyper parameters, like learning rate, generation size and the fitness function

Code Bullet how is the fitness calculated?

Code Bullet What parameter values did you end up using? I’m making something similar and having issues with getting the right hyper parameters.

Yeah I had them die. I did this in hopes that it would encourage them to avoid their own bodies

Smart.

Craptastic idk man, I wouldn't want a learning algorithm playing a video game as my screensaver. What happens when your screensaver becomes self aware amd tries to kill you??? 😂😂

Millad Bahrami o Fuk

You can set webpages as screensavers with wallpaper engine, if this runs on javascript in the browser it would be possible

.

This reply is just here so I get a notification if they respond

John Doe ahaha

you can play with my code :^) bjornsundin.com/projects/neural-networks-evolving-to-play-snake/index.html

thanks a lot : )