The difficult thing about this, is that when you die the blue dots don't reset with you, so that'll change the way the AI learns and would be very hard to implement.
2:45 Straight line doesn't matter, if and only if a part of your selection is big and random enough. Thus after enough tries a random one will be better then the straight line, if the advantage is big enough not to go straight in a line.
Can be done in the same way, Geometry Dash is hard for humans, due to precision timing required, but learning the winning sequence could be done using a very similar search-based technique as here. Creating an AI that could solve an unseen Geometry Dash level at first go would be way, way harder, and is closer to what the OpenAI, Deep Mind etc teams are going for.
Neil Slater I like to think that AIs would easily solve certain games after some trial and error just like The World's Hardest Game. Unfortunately Geometry Dash is very very complicated timing wise. To get the AI to learn precision timings is extremely difficult. But once the AI learns, it will never die to a part that it has already learned. Humans are the exact opposite as humans have a higher senses and can see the game as the way it is. The only thing is that humans aren't perfect. They lack the consistency of what they learned and occasionally mess up. It still is a great idea nonetheless.
The problem most people had with this on the first video and this one is that it isn’t really artificial intelligence it’s just trial and error. By definition it is AI however, all this is doing is just moving the square randomly, hundreds of times, until you get closer to the goal. It hasn’t learned anything, and it’s made no decisions. What people wanted was some sort of neural network that will learn all the game rules and then play by itself, one character with no trial and error, this would be much more entertaining and interesting.
i.e. some sort of NN with 8 sensors in all directions detecting wall and distance to it; 8 sensors detecting death and distance to it and 8 sensors detecting coin and distance to it or something And adjusting fitness accordingly(i.e. as a function of time taken, distance to finish and coins)
Can't wait for you to do more level. However, a little complaint...ish. This was said on the last video's comment section but it bothers me too so I'll repeat it. The AI is just looking for the perfect path instead of trying to avoid the dots. It's more of a guesswork game rather than brilliant AI. Otherwise though, amazing video as always.
They have done that. The AI didnt understand how to think ahead and just paused the game when it was about to loose, so it didnt. Google it if you dont believe me, i dont remember the name of the guy who did it, sorry.
"He must be so bored thatvhe has to resort to killing himself multiple times for entertainment; I mean, its relatable really, we've all been there" lmbo
What if you made a genetic algorithm that teaches two "players" to kill and shoot each other? I saw another video do it but i need a code bullet versionnnnn
I don't think using a plain genetic algorithm for this was a great idea since it only learns the path for that specific level. I think giving them all NN's would work better, then saving the parameters at the end of each level. That means instead of learning a specific path, it learns what to do on new unseen levels as well by learning importance on values such as distance to walls, blue dots and coins to apply on later levels.
Is it a reference only or is it a math problem I Could actually solve if so could you re write (2^18) in words? Edit: to solve it is it (2^18)=2•2•18=72 So 262,144 • 72 = 18 874,368??? Or is it 262 144??? I’m asking bc idk why but I like solving big and harder math tasks randomly like this but I hate math in school?!!😹👍🏻
You were meant to make these videos. thank you for you contribution to society, and a personal thanks for helping me grasp neural networks on another level. Seriously, I've been learning things from your videos that TED talks didn't present this way. Not to sound like a weird internet person LOL but it'd be awesome to nerd out with you on a project, because so few people share your kind of perspective or enthusiasm toward this knowledge but Me, and 200,000+ others, do. Keep making videos because they offer much and the AI tests are extremely interesting. Especially for people like me, where the internet is all i have to further my understanding of these topics. So on behalf of almost a quarter million people and myself, LOL - Congratulations to you and the Channel, and thank you for all your hard work.
Anyone else notice that each of the AI's doing "nothing" segments were their own individual complex patterns and rhythms? It was both beautiful and worrying all at once.
I program with Java & JavaScript myself and I love checking out the source code of your programs, will that still be available? I probably won’t get an answer, so I shouldn’t be trying... Love your videos by the way! I was actually thinking about creating coding videos soonish, but you beat me to it! (you’ve really changed since the first video, I’ve been here since the start) Keep up the great work! :D
usually i hate it when people say "ive been here since you had 7 subs," but since this guy said it, im gonna make it a competition. ive been here around about when you made... oh wait just checked, ive been here since the start too, xd. the snake ai didnt seem too long ago. damn this channel grew p quickly compared to others
Do you think it's possible to have them learn on randomly interchanging levels until they learn how to complete any level? Maybe by letting the ai know how far away they are from the dots?
He could probably train a neutral network into learning the "rules" of the game, rater than just using a genetic algorithm to find a series of steps that finishes a single level. A model of the rules of the game, may be able to complete any level that follows the same rules (assuming the level is solvable, naturally). That is way more difficult though. Training neural networks using genetic algorithms is a very different, albeit interesting, kind of beast.
In the gosu library (for ruby and C++) there is a distance function that takes the x and y of both points to give, well, the distance. I don't know about java though.
Its not even a good enough AI too do that (and doesn't have enough control) Programmers would stop an AI enslaving humanity before it happens with all sorts of restrictions, so it physically can't happen. An AI wouldn't even want to enslave humanity.... Lol
Is Javascript really that great for building AI? I understand he's just using JS to host his projects, but aren't languages like Java or C++ more flexible. Would you recommend Java as a starter
Dilan Omer First of all I am by no mean an AI expert. I would agree that C++, Ruby etc would be a lot more powerful because they are low level languages and you can utilize all cores. Java and C# would good too even though they run in a virtual machine. Soo why Javascript? Well i personally love javascript because it is extremely deployable, its fast to write code, its a multiparadigme language, tons of open source code and no compiling / no waiting. The only thing its really missing is to have strong types which Typescript is fixing, also immutable types by default would be nice too xD
If you want to make ai i suggest python, not only python is suggested as a beginner language but also is the most common language for ai, it is not really fast but you will most likely use it with tensorflow or something that is written is c++ and has a wrapper for python, so you can understand it easilly in python syntax.
Oh wow! You showed my fan-made worlds hardest game on scratch at first! I made that back in 2013 and at the time, I sucked at coding, but I've gotten a lot better over the years.
Love your videos. Actually I'm kind of proud to be part of your demographic stats in such a low category so early on on your channel. *cough* female and really like your content. Keep up the good work! It's inspiring.
I edited this comment because I was asking for the decoding the coding device video, but this video is fine. I should be more appreciative. Thank you for taking the time out of your day to make a video, codebullet.
Nico Gutholz it is the go to game engine for people who can't program for shit. Unity and Unreal is better, if you're really good... make yourself a in-house game engine. ;p
i think you are completely making a mistake about the purpose that the ai should learn. you should try to code random levels with blue balls going around randomly and with random location for the green spaces and the coins. THEN that your player learned to take all the coins and reach the green spaces dodging the blue dots you should test it on these levels of the hardest game. in your case you are just brute forcing each level
erre4 I don’t code but I don’t think you can take an A.I’s experience from a random level and put them into a level from the game. It’s how we learn pretty much we figure it out by ourselves
Halomaster 213 i code and this is really what an a.i. is. you train a model on some data, in my case the data is a large set of random levels, then the ai learns that it must dodge the blue dots take the coins and reach the green space. once that on most of your data your ai is able to win, you can say that your ai learned that it should dodge the blue dots and take the coins. in this case your a.i can play (if learned well) every game of the hardest game without retrain the model for each level, because it learned the general purpose. what c.b. does is just brute forcing some parameters for a determinate level, but that a.i. can't generalize for other levels indeed he retrained for different levels, and this is not an a.i.
His WHG AIs so far are the equivalent of putting a blind man in an obstacle course without any instructions, and waiting for him to finish it by accident. He's already done more intelligent AI using NEAT, such as the ones for Snake and Asteroids. Employing that same model here would be a lot more interesting, and could perhaps learn how to solve levels it's never encountered before on the first try
It's also technically not bruteforcing as it is not randomly trying every possible movement pattern, as that would take years, but instead evolving its movement pattern based on a set goal. A genetic algorithm is still machine learning, which is what many consider an AI. I personally know a lot of people dislike the term AI as it gives false implications on what machine learning really does. There are more advanced machine learning algorithms like what you are asking for , and I would love to see them, however they are extremely complicated as well. And applying them to beat this first try would be far more complicated then how he used NEAT to learn how to play asteroids and require lots more data.
Dirty Poul he can find one, it’s called a neural network. Lots of the people commented on his last video saying use a NN instead of this poor joke at an AI that just uses trial and error. But he ignored them because he’s too lazy.
This AI just moves randomly, and whatever sequence gets closest to the target carries on to the next generation. It hasn’t learnt how to play, it’s just following a sequence of random moves.
You are a fucking beast! Thanks for sharing all of this stuff, it's so entertaining. You put so much effort, prepare it so nicely for us, thank you mate, seriously, we appreciate it!
Looking through your code in github this seems just an algo that finds a combination of movements to solve the level blindfold. Change a dot position by the sligthleast and it will fail. It just finds a succession of keyboard inputs so to speak. This is not AI. There is zero perception of the environment and zero abstraction of what it takes to dodge a blue dot.
Well, yeah, it's just a genetic algorithm, and it works only for one level without relearning the whole thing, and it's the point. Although I do agree it's inaffective.
True while it could exist in a way that allows it to retain knowledge accumulated through levels it is a program which learns even if it's through brute force meaning it does definitely qualify as ai
It's not that hard to go from one programming language to another. You know what you want to code, you just need to google the syntax for that other language. And a lot of basics are identical in most of the languages.
modern functional javascript is nice. you can go pretty far while ignoring all the old java-inspired crud. destructuring and higher order functions are just so much fun. :D
I don't think this is an AI. An AI would try to recognize the pattern and not just blindly try every possibility. This solution would immediately break, if you randomize the starting condition.
I can write a program to brute force a password. That program generates combinations that can solve the password, but it is not an AI either. Evolution isn't AI.
The AI that solves this puzzle, could technically also solve a puzzle that is similar, but not the same. You could make 100 different puzzles and evolve an agent that can solve all of them. It makes choices based on where it is at any time. The choices it makes isn't finite, there are infinitely many input possibilities it has to distinguish between. It is not the same as a combination, or a password. It's not just a key that solves THIS puzzle, it's a general purpose agent that can solve any puzzle of this kind. It isn't a sequence of moves. It's a logic tree that chooses moves based on where it is. Actually, that's what code bullet should have done. He should have had the fitness function test more than one map. To see how well it solves a puzzle of this kind. Rather than this specific puzzle.
also idea, leauge AI? (pretty much just learning how to win, also might take like 3 months through raw gameplay and more trying out different combinations of champions lol)
Everyone in the comments like "this isn't an AI it's brute force blah blah blah make a real AI" like why don't you do it yourself huh? This reminds me of when children tell me to solve a rubik's cube faster. Like unless you can solve a rubik's cube, you can't tell me how fast I'm supposed to be
I totally agree but it's not about intelligence, just learning. For me this AI is like a blind player that only knows when his dying. He don't know where the wall/blue dot are. So the player just remember when his dying and what key combination he did. The AI here just remember Left - Right - Left - Down until it works, where a normal player thinking is more like Left to avoid this dot then Right to avoid this one. The only difference is that the AI only use direction where the player use direction by taking in consideration wall and blue dot.
Only if it allows it to then perform similar tasks through what it's learned. With this program, if you start it from a slightly different position on the exact same level, it has to 'learn' from scratch again. IE, it hasn't really learned anything, it's just coincidentally found a winning set of inputs that will only work under those exact starting conditions. To learn is to generalise (apply what's been learned to other situations). I recommend looking up the term 'overfitting machine learning', as that describes what his program has done.
I read your comment and I totally agree. Interestingly enough though psychology's definition of learning (from what I've learnt) is "constant change in behavior that is formed due to previous experience", which means that this ai is learning (according to psychology)...
This is hilarious. I am hoping to see them beat all the levels. Get better soon! Edit-I remember when this was on cool math games and you could go to any level. Ah my middle school days...
Dinosorusmechkiller See how he completed the levels faster than the AI. He learned how to beat the obstacles, the AI randomly moved and the best one was shown on screen.
This is not AI, it's just brute force. For it to be AI it needs to learn to avoid the dots, so when you get to the next level, it should have some idea what to do.
He means that the AI, after having learned how to beat one level, still has no idea how to solve the next and needs to go through the process again for the next level, because it doesn't learn to avoid the blue dots, it just learns to get as close as possible before being brutally murdered.
I think you can build a more general AI by feeding it with huge amount of game variation and ask it to solve them, then it may learn how to recognise familiar game patterns in an entirely new game setting. But I guess it will require a much more sophisticated deep learning programme. Somewhat similar to AlphaGo/AlphaZero.
YipHyGaming - Truncation [150 coming] i mean, it is trying every move possible, 700 'players' trying different moves and then using the one that got the furthest to try 700 more.. not sure i see the difference, or the need for such hostility on a fun video
CraftyBLUEstone he would probably code it with a similar AI to the dinosaur game, although geometry dash is way more complex to code. He likes to code simple games to see if an AI can beat it, but I think an AI could beat geometry dash.
I think I like to watch the entire population all try for it at the same time. It might be a bit more confusing, but at the same time I find it almost mesmerizing.
Not sure this can be classed as "AI". Still pretty neat but, it's just making small changes randomly. AI needs to be able to perceive patterns and rules, and use this knowledge to complete a task.
In scratch hit detection is so that it forces you to use detection based contact with a specific color color in order to be accurate, in the barriers case, black
Make a AI which creates more levels for the AI ;)
That's a good way to get a self aware AI
Sounds like a plan
LJSystems he better do that
Bob Harrison no it doesn't
Bob Harrison it can't become self aware on it's own unless it Is programmed to do that, and believe me, it isn't easy
The difficult thing about this, is that when you die the blue dots don't reset with you, so that'll change the way the AI learns and would be very hard to implement.
E
Time for a confession.
I don't understand ANY of the technical terms.
I'm just here for the results.
me too
lol same
Who isn’t lol
I’m just here because he is super funny
We all are friend
2:45 Straight line doesn't matter, if and only if a part of your selection is big and random enough. Thus after enough tries a random one will be better then the straight line, if the advantage is big enough not to go straight in a line.
"We reached 200000 subscribers!" 2 seconds later: 1.5 million subs
4 seconds later: 2.5 million subs
@@UltraVioletLite yes
Make an AI that can play Geometry Dash.
Imagine it beat Crimson Planet or Bloodlust or something xD
Can be done in the same way, Geometry Dash is hard for humans, due to precision timing required, but learning the winning sequence could be done using a very similar search-based technique as here.
Creating an AI that could solve an unseen Geometry Dash level at first go would be way, way harder, and is closer to what the OpenAI, Deep Mind etc teams are going for.
Neil Slater I like to think that AIs would easily solve certain games after some trial and error just like The World's Hardest Game. Unfortunately Geometry Dash is very very complicated timing wise. To get the AI to learn precision timings is extremely difficult. But once the AI learns, it will never die to a part that it has already learned. Humans are the exact opposite as humans have a higher senses and can see the game as the way it is. The only thing is that humans aren't perfect. They lack the consistency of what they learned and occasionally mess up. It still is a great idea nonetheless.
Jazzling I know right. It'll probably become the new auto-clicker but you don't have to program it.
Kim Nasayao What a great idea
Finally, a comment section without Justin Y.
Edit: He found us...
Lol
*knocks on wood furiously
He’s here watching
you under estimate his powers. his recommended is unlike anything you’ve seen.
Dont jinx it
The problem most people had with this on the first video and this one is that it isn’t really artificial intelligence it’s just trial and error. By definition it is AI however, all this is doing is just moving the square randomly, hundreds of times, until you get closer to the goal. It hasn’t learned anything, and it’s made no decisions.
What people wanted was some sort of neural network that will learn all the game rules and then play by itself, one character with no trial and error, this would be much more entertaining and interesting.
Lodlum B I agree, I‘d love to see an AI, that „knows“ the rules and can play any level @CodeBullet
Lodlum B Finally, a voice of rationality amongst a sea of “MOAR LEVELS”.
Thanks for reading
i.e. some sort of NN with 8 sensors in all directions detecting wall and distance to it; 8 sensors detecting death and distance to it and 8 sensors detecting coin and distance to it or something
And adjusting fitness accordingly(i.e. as a function of time taken, distance to finish and coins)
It's not an AI, it's a Genetic algorithm for vectors.
I just ran you AI and it is truly beautiful even your new work this still holds up and is such a joy to watch it
I love this type of content. Keep going my friend!
Can't wait for you to do more level. However, a little complaint...ish. This was said on the last video's comment section but it bothers me too so I'll repeat it.
The AI is just looking for the perfect path instead of trying to avoid the dots. It's more of a guesswork game rather than brilliant AI. Otherwise though, amazing video as always.
Why do Java programers always wear glasses?
Because they can't C#
All hail the glory of the Nerd Pun!
That just may be the best pun i've ever seen :D
no smh just no
nice
unoriginal
MOAR!!!!!!!!!!
Andrew W agreed
Go away
Morreeeeeeeeeeeee!!!!!!!
Mooar
AI learns to play tetris
THEFOXDEMON21 YES
*Boom! Tetris for AI*
@@shovelozubejtygr nah, the AI can never cause as much hype as Jeff
BOOM! TETRIS FOR JEFF!
They have done that. The AI didnt understand how to think ahead and just paused the game when it was about to loose, so it didnt. Google it if you dont believe me, i dont remember the name of the guy who did it, sorry.
AI learns tetris 99
"He must be so bored thatvhe has to resort to killing himself multiple times for entertainment; I mean, its relatable really, we've all been there" lmbo
3:52 A meme review fan, of course, it is the most entertaining show in the internet
youtube favorite show
Pichaman UA-cam’s favorite show of all time
Thats why i went to the comments lmao
meme review is a meme killer
👏👏
There are WAY more levels
Yeah booiiiiiiiii
I think he knows.
Sigh people that don't watch the whole video
What if you made a genetic algorithm that teaches two "players" to kill and shoot each other? I saw another video do it but i need a code bullet versionnnnn
Definitely this!
do you mean Albert ?
Make one AI that can play all levels rather than one specialized AI for each level
I love to see other people who like AI! I make videos about it, let me know what you think!
@@PandoraMakesGames ua-cam.com/video/JwncAQurlqk/v-deo.html
I don't need to explain.
Yeah, because it's that simple
I don't think using a plain genetic algorithm for this was a great idea since it only learns the path for that specific level. I think giving them all NN's would work better, then saving the parameters at the end of each level. That means instead of learning a specific path, it learns what to do on new unseen levels as well by learning importance on values such as distance to walls, blue dots and coins to apply on later levels.
@@ardaozden887 it would also take way too long
Code Buller: I'm actually in the early stages of being a bit sick.
2020 Viewers: REEEEEEEEEEEEEEEEEEEEEEEEEEEEEEE!!!!
No! The next milestone is 262,144 (2^18)
Gal Horowitz
3B1B references. -3- boi im watching you
shhh
Is it a reference only or is it a math problem I Could actually solve if so could you re write (2^18) in words?
Edit: to solve it is it (2^18)=2•2•18=72
So 262,144 • 72 = 18 874,368???
Or is it 262 144???
I’m asking bc idk why but I like solving big and harder math tasks randomly like this but I hate math in school?!!😹👍🏻
2^18 means two to the power of eighteen...
Lucas Svedlund
All you have to do is 2×2 and make sure theres 18 2's
More levels!!!!!!!!!
Petar Vrdoljak znas marina?
Kojeg???????
*click*
this channel is the wae
do Tetris!
Taylor youngreen yes do tetris!
Jasprr_Yeet of course it can xD
A convolutional neural network could solve Tetris. I doubt a NEAT neural network could, though.
That's actually a pretty good idea.
TETRIS EFFECT
You were meant to make these videos. thank you for you contribution to society, and a personal thanks for helping me grasp neural networks on another level. Seriously, I've been learning things from your videos that TED talks didn't present this way. Not to sound like a weird internet person LOL but it'd be awesome to nerd out with you on a project, because so few people share your kind of perspective or enthusiasm toward this knowledge but Me, and 200,000+ others, do. Keep making videos because they offer much and the AI tests are extremely interesting. Especially for people like me, where the internet is all i have to further my understanding of these topics.
So on behalf of almost a quarter million people and myself, LOL - Congratulations to you and the Channel, and thank you for all your hard work.
Anyone else notice that each of the AI's doing "nothing" segments were their own individual complex patterns and rhythms? It was both beautiful and worrying all at once.
I program with Java & JavaScript myself and I love checking out the source code of your programs, will that still be available?
I probably won’t get an answer, so I shouldn’t be trying...
Love your videos by the way! I was actually thinking about creating coding videos soonish, but you beat me to it! (you’ve really changed since the first video, I’ve been here since the start) Keep up the great work! :D
Thx, also, Code Bullet loved my comment? WOW! Didn’t expect that! This really made my night!
Anthony Willson: notice meh code senpai!!! Lol
Your comment got loved, I’m proud of you!
usually i hate it when people say "ive been here since you had 7 subs," but since this guy said it, im gonna make it a competition. ive been here around about when you made... oh wait just checked, ive been here since the start too, xd. the snake ai didnt seem too long ago. damn this channel grew p quickly compared to others
Nice job man!
Do you think it's possible to have them learn on randomly interchanging levels until they learn how to complete any level? Maybe by letting the ai know how far away they are from the dots?
Lion's Tail this would be super cool!
He could probably train a neutral network into learning the "rules" of the game, rater than just using a genetic algorithm to find a series of steps that finishes a single level. A model of the rules of the game, may be able to complete any level that follows the same rules (assuming the level is solvable, naturally). That is way more difficult though. Training neural networks using genetic algorithms is a very different, albeit interesting, kind of beast.
Not with what he's doing here. But theoretically, yeah that seems like it aught to be possible.
In the gosu library (for ruby and C++) there is a distance function that takes the x and y of both points to give, well, the distance. I don't know about java though.
you could use NEAT for that, right?
I'm watching your videos a while now, and I think it's now the time to tell you: I f'ing love them! Please continue so !
Following you from Italy, your videos are really great
MORE LEVELS!!!
The AI moves slow....
In the atcual game it's quite quicker
Aaaaand I'm still waiting for the enigma machine pt.2!!!
This has been my new favorites ever channel to get notifications from! Damn good contents.
So happy your channel is working so well now
Yay finally more levels :D
Hmmm I challenge you to make an Evolution making an Evolution
MOAR
(plz)
Getting sick in:
2018: Calm
2020: *PANIK*
This channel is going places! Congrats!
So when does this AI enslave humanity exactly?
Anthony Smith the program would crash before that
Its not even a good enough AI too do that (and doesn't have enough control) Programmers would stop an AI enslaving humanity before it happens with all sorts of restrictions, so it physically can't happen. An AI wouldn't even want to enslave humanity.... Lol
wooosh
This is precisely what I wanted to know. It is not a wooosh moment. :3
T o t a l l y
200,000
Then I realize how long I’ve been putting of watching this video lmao
What languages do you code your projects in? Where did you learn to build AI? Great videos btw
Dilan Omer Java in processing, but he is using javascript now. If you to learn making AI i suggest look at coding trains AI and intelligence series
Is Javascript really that great for building AI? I understand he's just using JS to host his projects, but aren't languages like Java or C++ more flexible. Would you recommend Java as a starter
Dilan Omer First of all I am by no mean an AI expert.
I would agree that C++, Ruby etc would be a lot more powerful because they are low level languages and you can utilize all cores. Java and C# would good too even though they run in a virtual machine.
Soo why Javascript? Well i personally love javascript because it is extremely deployable, its fast to write code, its a multiparadigme language, tons of open source code and no compiling / no waiting. The only thing its really missing is to have strong types which Typescript is fixing, also immutable types by default would be nice too xD
+Dilan Omer He built in javascript so that we could experiment with it on his website. He normally uses Java.
If you want to make ai i suggest python, not only python is suggested as a beginner language but also is the most common language for ai, it is not really fast but you will most likely use it with tensorflow or something that is written is c++ and has a wrapper for python, so you can understand it easilly in python syntax.
Oh wow! You showed my fan-made worlds hardest game on scratch at first! I made that back in 2013 and at the time, I sucked at coding, but I've gotten a lot better over the years.
Congratulasion on 200 subs dude! Big fan
-3:51-- Level Review?-
Careful when you're asleep next time your AI will conquer the world
Do another game! xD
Love your videos. Actually I'm kind of proud to be part of your demographic stats in such a low category so early on on your channel. *cough* female and really like your content. Keep up the good work! It's inspiring.
keep on making great videos Mr Memio
Edit: got a like so someone got the reference
MORE LEVELSSSSS
Great job by the way! Congratulations! It just took one sanity to achieve it!
"He must be so bored to have to resort to killing himself repeatedly for entertainment"
Code Bullet's AI casually predicting quarantine.
lol
Build an A.I. that uses paint or sketch programs to generate art from imputs that have been collected from other UA-cam pages.
Ai learning plaigurism.
I think people want more levels
Nice try to get into the next video
You can use javascript in Unity, And compile the project for web :)
I edited this comment because I was asking for the decoding the coding device video, but this video is fine. I should be more appreciative. Thank you for taking the time out of your day to make a video, codebullet.
Just to note that the generation number and the number of moves on the second level spell out 1 2 3 4 5. Thank you and good night
It’s said Code Bullet is still trying to beat the levels he already did so he can make the next video.
He could just watch the speed runs you know
why are you playing on scratch?
OH SHIT U LETTING US RUN THE AI NICE
Scratch is the superior game engine
Nico Gutholz it is the go to game engine for people who can't program for shit.
Unity and Unreal is better, if you're really good... make yourself a in-house game engine. ;p
飛鳥「月」 I was just kidding
boredbenny Because the game was made on scratch
AI learns to play the scary maze game
TheElement OfFyre
But... But does it punch?
thats what i thought the original was when i saw the thumbnail =)
Second it
I love watching the ways AI play that a human would never do. The straight line to the coin approach gets such interesting results
im glad that you made you programs so accessible, youre awesome dude
do a flappy bird ai
and snakes
up
Qz Xyst down
Left
Right
He has already made a snake AI.
i think you are completely making a mistake about the purpose that the ai should learn. you should try to code random levels with blue balls going around randomly and with random location for the green spaces and the coins. THEN that your player learned to take all the coins and reach the green spaces dodging the blue dots you should test it on these levels of the hardest game. in your case you are just brute forcing each level
erre4 I don’t code but I don’t think you can take an A.I’s experience from a random level and put them into a level from the game. It’s how we learn pretty much we figure it out by ourselves
Halomaster 213 i code and this is really what an a.i. is. you train a model on some data, in my case the data is a large set of random levels, then the ai learns that it must dodge the blue dots take the coins and reach the green space. once that on most of your data your ai is able to win, you can say that your ai learned that it should dodge the blue dots and take the coins. in this case your a.i can play (if learned well) every game of the hardest game without retrain the model for each level, because it learned the general purpose. what c.b. does is just brute forcing some parameters for a determinate level, but that a.i. can't generalize for other levels indeed he retrained for different levels, and this is not an a.i.
His WHG AIs so far are the equivalent of putting a blind man in an obstacle course without any instructions, and waiting for him to finish it by accident. He's already done more intelligent AI using NEAT, such as the ones for Snake and Asteroids. Employing that same model here would be a lot more interesting, and could perhaps learn how to solve levels it's never encountered before on the first try
It's also technically not bruteforcing as it is not randomly trying every possible movement pattern, as that would take years, but instead evolving its movement pattern based on a set goal. A genetic algorithm is still machine learning, which is what many consider an AI. I personally know a lot of people dislike the term AI as it gives false implications on what machine learning really does. There are more advanced machine learning algorithms like what you are asking for , and I would love to see them, however they are extremely complicated as well. And applying them to beat this first try would be far more complicated then how he used NEAT to learn how to play asteroids and require lots more data.
Soooooo Asteroids? Which he already did?
You should use the same generations for all levels, so you can get a pro AI that can beat all levels.
AlexINF Yea I was thinking the same, make a fucking AI that beats the fucking game
It'll probably need to forget the last level to relearn the next level. But who knows?
MikeFPV You're right, he'd need a different approach. But if he finds one it would be much more interesting.
Dirty Poul he can find one, it’s called a neural network. Lots of the people commented on his last video saying use a NN instead of this poor joke at an AI that just uses trial and error. But he ignored them because he’s too lazy.
This AI just moves randomly, and whatever sequence gets closest to the target carries on to the next generation. It hasn’t learnt how to play, it’s just following a sequence of random
moves.
You are a fucking beast! Thanks for sharing all of this stuff, it's so entertaining.
You put so much effort, prepare it so nicely for us, thank you mate, seriously, we appreciate it!
Just wanted to say thanks for the videos and love everyone of them.
Do the other levels plz
Your love of pewdiepie gives me joy!
Looking through your code in github this seems just an algo that finds a combination of movements to solve the level blindfold. Change a dot position by the sligthleast and it will fail. It just finds a succession of keyboard inputs so to speak. This is not AI. There is zero perception of the environment and zero abstraction of what it takes to dodge a blue dot.
You are absolutely right about this! I make AI videos, let me know what you think!
Shameful promotion much?
Well, yeah, it's just a genetic algorithm, and it works only for one level without relearning the whole thing, and it's the point. Although I do agree it's inaffective.
I'd argue this is a distinction between AI in the game context, and AI in a more technical context.
This absolutely qualifies as AI in the game sense.
True while it could exist in a way that allows it to retain knowledge accumulated through levels it is a program which learns even if it's through brute force meaning it does definitely qualify as ai
Wow you were busy! Excellent work and congrats on 200K :O
Wonder how much frustration that AI is going through doing this
Hey CB! Could you make a tetris AI? That would be amazing
How exactly does one "pick up" javascript. What kind of superhuman are you.
It's not that hard to go from one programming language to another. You know what you want to code, you just need to google the syntax for that other language. And a lot of basics are identical in most of the languages.
I'm fluent in a couple of languages, I understand, but javascript. JAVASCRIPT.
Javascript is not that bad. Better than PHP at least.
for games ?! not realy sure...
modern functional javascript is nice. you can go pretty far while ignoring all the old java-inspired crud. destructuring and higher order functions are just so much fun. :D
I don't think this is an AI. An AI would try to recognize the pattern and not just blindly try every possibility. This solution would immediately break, if you randomize the starting condition.
P8qzxnxfP85xZ2H3wDRV this is an ai that uses evolution.
it tries someting it fails so it tries again and again wit changing the last part only
Survival of the best is not intelligence.
It's not the instance of the player that solves the game that has AI, it's the program generating players that can solve the game, that is AI.
I can write a program to brute force a password. That program generates combinations that can solve the password, but it is not an AI either.
Evolution isn't AI.
The AI that solves this puzzle, could technically also solve a puzzle that is similar, but not the same.
You could make 100 different puzzles and evolve an agent that can solve all of them.
It makes choices based on where it is at any time.
The choices it makes isn't finite, there are infinitely many input possibilities it has to distinguish between.
It is not the same as a combination, or a password. It's not just a key that solves THIS puzzle, it's a general purpose agent that can solve any puzzle of this kind. It isn't a sequence of moves. It's a logic tree that chooses moves based on where it is.
Actually, that's what code bullet should have done. He should have had the fitness function test more than one map. To see how well it solves a puzzle of this kind. Rather than this specific puzzle.
45k subs in 5 days, insanity. also, nice work and more levels
also idea, leauge AI?
(pretty much just learning how to win, also might take like 3 months through raw gameplay and more trying out different combinations of champions lol)
Everyone in the comments like "this isn't an AI it's brute force blah blah blah make a real AI" like why don't you do it yourself huh?
This reminds me of when children tell me to solve a rubik's cube faster. Like unless you can solve a rubik's cube, you can't tell me how fast I'm supposed to be
I'd like to see all the levels XD
I think the ai just remembered the path and didn't learned to play it
Yes exactly. But isn't that the definition of learning ? Remembering stuff.
I totally agree but it's not about intelligence, just learning. For me this AI is like a blind player that only knows when his dying. He don't know where the wall/blue dot are. So the player just remember when his dying and what key combination he did. The AI here just remember Left - Right - Left - Down until it works, where a normal player thinking is more like Left to avoid this dot then Right to avoid this one.
The only difference is that the AI only use direction where the player use direction by taking in consideration wall and blue dot.
It did learn the path via a brute force algorithm, which makes it more a machine learning thing than artificial intelligence
Ai learns to play Tetris, plz
See that would require him to actually make a program that can *learn,* unlike what he's done in this video which is just trial and error.
He has made ones that learn in other videos, like the snake one. Also I hate phone keyboards.
A Rocket Trial and error is still a form of learning, isn't it?
Only if it allows it to then perform similar tasks through what it's learned. With this program, if you start it from a slightly different position on the exact same level, it has to 'learn' from scratch again. IE, it hasn't really learned anything, it's just coincidentally found a winning set of inputs that will only work under those exact starting conditions. To learn is to generalise (apply what's been learned to other situations).
I recommend looking up the term 'overfitting machine learning', as that describes what his program has done.
I read your comment and I totally agree. Interestingly enough though psychology's definition of learning (from what I've learnt) is "constant change in behavior that is formed due to previous experience", which means that this ai is learning (according to psychology)...
This is hilarious. I am hoping to see them beat all the levels. Get better soon!
Edit-I remember when this was on cool math games and you could go to any level. Ah my middle school days...
More levels definitely. You’re dope.
I wish i was just as good as that Ai.
Dinosorusmechkiller See how he completed the levels faster than the AI. He learned how to beat the obstacles, the AI randomly moved and the best one was shown on screen.
This is not AI, it's just brute force. For it to be AI it needs to learn to avoid the dots, so when you get to the next level, it should have some idea what to do.
I want to learn more please explain
He means that the AI, after having learned how to beat one level, still has no idea how to solve the next and needs to go through the process again for the next level, because it doesn't learn to avoid the blue dots, it just learns to get as close as possible before being brutally murdered.
I think you can build a more general AI by feeding it with huge amount of game variation and ask it to solve them, then it may learn how to recognise familiar game patterns in an entirely new game setting. But I guess it will require a much more sophisticated deep learning programme. Somewhat similar to AlphaGo/AlphaZero.
Like in the first attemps learn the patrons of movement of the dots, then trying to avoid then while moving to the goal
YipHyGaming - Truncation [150 coming] i mean, it is trying every move possible, 700 'players' trying different moves and then using the one that got the furthest to try 700 more.. not sure i see the difference, or the need for such hostility on a fun video
Thank you!
Yes, more. Please. Also, I like that idea of this being a side series.
5:15 - So for your "move" options, is "wait" one of them, or just directions?
Create AI that makes random games to play
More levels Kappa
earth is a donut fine 😡
What does kappa mean
earth is a donut thank you
Aerede krappa. More than kappa
earth is a donut
that’s no what Kappa mean you fucking normie DansGame
A little disappointed that I didn't find this channel earlier
Ayy you hit 200 sub
Congrats xD
The entire game. You can even do it over multiple videos, I'll watch each of every one of them.
Can AI play geometry dash?
CraftyBLUEstone he would probably code it with a similar AI to the dinosaur game, although geometry dash is way more complex to code. He likes to code simple games to see if an AI can beat it, but I think an AI could beat geometry dash.
TWO HUNDRED!!!!
Notification squad
Ur name should be "no internet dinosaur"
Twitter Squad
Oh no, the wifiless monster
Orichishu the hell's your problem.
Damn you must realy have problems with Internet.
”Im in the early stages of being a bit sick” every one in 2020: ”Corona viruuus!!”
I think I like to watch the entire population all try for it at the same time. It might be a bit more confusing, but at the same time I find it almost mesmerizing.
This is how you get GLaDOS
this.videoLiked = true;
More this.()
always more this.()
Not sure this can be classed as "AI". Still pretty neat but, it's just making small changes randomly. AI needs to be able to perceive patterns and rules, and use this knowledge to complete a task.
agreed
it needs to know the game and the rules not the level
but its steal learning, thats the peice of ze rezizstonceokcSELPFHSltnS
In scratch hit detection is so that it forces you to use detection based contact with a specific color color in order to be accurate, in the barriers case, black
2:55, cool I like it more if an AI is learning on it's own than if it gets a headstart