/v/ - Video Games

Are you going to give us a link to the video or the stream or whatever?

>>14048521

>>14048554

https://www.youtube.com/watch?v=-awB6mPlRag

Use the direct link or use streamlink to watch it. It's a livestream.

>>14048563

I know, I just thought that was retarded.

>>14048507

if it's running on normal speed it isn't too surprising

if the network wasn't trained on stairs that is exactly what you would expect.

its the researchers fault for not introducing proper training.

>>14048736

The entire point of this particular AI is that it learns by itself.

YOU CAN DO IT AI CHAN

>>14048507

It's a neural network specifically for learning mario games and it isn't provided with any information aside from surface, moving object (ALL moving objects identified as identical entities) and distance travelled right.

How is this specific to video games or mario platformers especially if the only thing it does is learn to press buttons?

>Still hasn't gone over the last flight of stairs

Why couldn't he just train the AI to press jump on the flight of stairs instead of trying to self teach it?

Hehe, one "something" boi is currently my favourite meme. *3*

He should probably stop restarting the level and just let the AI figure it out.

>>14050871

Faggot, never have children if you're going to act like this towards them

Has this really been stuck on the stairs for 9 days? At least there's good music on the stream.

>>14050883

I'm going to gently enter your rear with my penis.

>>14050969

>being gentle

You some sort of soft faggot?

>>14048507

>>14048736

Reward shaping could fix the stair problem if this is a reinforcement learning model. Another thing would be an AI "imagining" a portion of the screen in a virtual space (to simplify probabilities) and project a group of actions to find the most probable reward success and then try it live. If a sequence works, it remembers it and reapplies that behavior when confronted with a similar situation. Eventually it'll learn how to stack a sequence when confronting stairs and get over them.

>>14051626

I read through his neural network programming, it's based on evolution through reinforcement. I don't know if he's using the same model he used in Super Mario World, but it's basically

>In game world is rendered to AI as a series of three blocks, white, gray, and black

>white space is surfaces such as floors, walls, platforms etc.

>gray is just open space

>black is hostiles

>"Fitness" scoring is the length Mario moves in a run (genome) and the time it takes to reach the end point

>After 100 genomes, a second species is ran with another 100 genomes

>After the second genome, the highest fitness genomes are selected as donor networks, that are given randomized variables making the next generation of tests

>Next generation ran, repeating the same format over and over ad-infinitum

What I do find of strange is that according to his videos, the AI is able to complete a Mario level in a day or two of learning, and a Mario Kart level within 6 hours. However, the AI hasn't gotten the neural network link figured out yet for the stairs that it could jump them even after this long. It can do the rest of the level almost perfectly including that almost frame perfect pit dodge jump.

>>14048507

The A.I everntually.

>>14048771

Hand it to the jewgle robot that plays chess. It would probably just copy a TAS input file to play through it.

>>14051882

If something unexpected happens, like needing to jump, it takes a long time to go through every single iteration of possibilities to finally conclude it is missing data that needs to be created. Which is to say trying to jump.

>>14051882

Probably could be improved by generating imaginary lines that project possible paths, and then it learns to interpret and create new line projections with corresponding actions. Stairs would generate a diagonal line at first in this case, however it would have already learned jumping looks like arcs by the time it reaches them. So eventually it'll learn to approximate multiple arcs to get close to the diagonal imaginary line, based on peak and valley points of each jump. Could result in much faster learning.

Human sort of do this too but it's more sophisticated, like an invisible abstraction of anticipated future movements or positioning of objects. You'll especially notice this "visual" sense when playing RTS games.

The reason it's been taking 9 days is because it LOSES fitness when it successfully reaches and enters the pipe. As such, dropping off the edge before the pipe can actually give more fitness than correctly taking the pipe. It still probably will eventually succeed, but because of that flaw in the measurement of fitness it's going to take forever.

>>14052037

I don't think that's the case, since it would probably either jump over the pipe in the first place or it would actively try to avoid getting hit on the turtles if that's how it was.

What's probably the problem is that when it gets deeper into generations the algorithm changes (in practice) from recognizing the threat/obstacle and avoiding it, to making mutations to recorded action sequences. The bigger the sequence, the bigger the consequences of the changes early on. But since the stairs are the last obstacle, several consecutive changes need to occur literally at the end of the sequence. There is also a possibility of the said sequences learning to be reluctant to change in general making the action required to be executed even less probable.

What probably should be done there, if a save load state is allowed and each stage is treated as a separate stage, then ai should recognize milestones of fitness it achieves thus creating sub, I dunno, generations, that try to continue advancing from said milestones which would make for smaller action sequences that require change in exchange for more intricate programming. For that purpose it should also probably recognize the fastest fpf it could achieve, but the problem here programmers aren't interested in the ai beating the games, that's why such changes wouldn't be made because they in a sense are an afterthought a crutch that devalues the original algorithm.

Also, the guy that is running this is probably not even related to the original creators of the algorithm and I don't think he would be doing any changes to the code since he only executed it and left it running just for the sake of it.

Why does it lose fitness when it jumps? Why would you want a Mario AI to not jump?

>>14048535

>>14048563

Y'know this is what we have Hooktube for right

>>14052763

Embed related

>>14052784

At least now we have evidence hooktube pushers are just a little bit retarded.

>>14052794

Just how do you think hooktube works? It just fetches the video files.

Is this already some sort of normalfag thing like salty or twitchplaysthing? Doesn't seem interactive enough for that. I am interested in MarI/O though.

>>14052768

Still counts as a view to YouTube AND you give your data to some random guy with no reason to trust him.

Where's the FOSS alternative?

>its on level 1-2

>it seems to know it has to be small mario to get by

>tfw it will never learn to just take the top path

>>14048507

When will it be able to do it without A presses?

>>14053101

Can AIs become autistic?

don't worry, in about 25 years this AI will be the world's best mario speedrunner known to man.

>>14053144

Is that not what the 'A' in AI stands for?

>>14053047

I don't get it, how many iterations does it need before it decides to jump? There are only two things you can do in SMB, it's run or jump, what the fuck is it waiting for? I blame the programmer.

>>14053174

I don't know much about ML but it looks like it is basically brute forcing the solution. Got very bored of watching when I figured that out.

>>14053174

The evolution goes backward from time to time. Species 2 forgot how to wall jump.

>>14053176

Yes, exactly. How about it jumps? It has already tried running into the wall all those fucking iterations.

I wonder if the ai will ever beat the game

>>14053201

Give 10 years min

>>14052763

>Why does it lose fitness when it jumps?

Imagine a line between Mario and the goal serving as the actual fitness measurement. By jumping, the distance between them momentarily increases as the goal is at ground level. Come to think of it, that would be the primary cause of the stairs problem if that's the case. Being so close to the finish makes it very unfavorable to consider jumping due to that.

>eugenics is ok as long as it's an AI

why are humans so hypocritical?

>>14053584

You can't apply eugenics to A.I

Just in case anyone doesn't understand, the fitness requirements are speed and moving to the right.

Hold right to win.

>>14053639

so it's meant for playing Sonic

What's with the technological troglodytes refusing to just open a link in another window? Is that too complicated for the average /v/ user now?

>>14053646

It would probably take a year for it to pass the first level of Sonic & Knuckles.

>It just gets stuck in various places

>Doesn't choose route that got best fit level

What a pointless ai. It can be improved to actually learn from its mistakes instead of trying every solution that just takes hours before it fucking jumps up the stairs.

I think the guy should check the microcircuitry of the AI. It was originally made for Super Mario World, and it obviously runs into some bug in Bros.

>>14053684

https://pastebin.com/ZZmSNaHX

Source code here

chat is getting niggered

>>14053782

To be honest this ai feels blind. How can you create a neural network if you can't create eyes for it to see impassable objects or objects that move? Its not real ai then, but more like a thing that remembers every trial and error and then combines all tries together in total darkness, failing to find the exit again and again. Even in fucking robots back in the 90's people were making means to make them see their surroundings to successfully navigate. You can't really tell that ai remembering its every trial and error truly evolves, because it never creates means for itself to adapt, like creating vision or senses for the next generation to survive successful.

It just remembers it progressed at certain point because it pressed jump in the dark.

>>14053817

They were asking for it.

>>14053047

it would be nice if we could see the inputs

it's possible that they might not be relative to mario's position, but just whatever is on the screen

Would be inserting to see this A.I play other games. I mean if it can't handle Super Mario Bros it would really blow a fuse over Mega Man 2.

>>14053820

It can "technically" see. It has white objects which are objects the player character can stand on and black objects that are moving objects. The only real problem I've actually noticed is the fact that the basis of success for the AI is Fitness. The higher fitness it averages, the more the AI is going to try and repeat the process.

When Mario falls in the same pit which causes death, but the total fitness increases as well, it causes complications where the AI itself thinks it might possibly be correct and continuously tries different inputs to get even farther even though it's a fault in coding. Basically the AI has more success getting to the pit and averaging out a higher fitness falling into the pit than jumping over it therefore it's stopping the AI from actually focusing on the stairs at the end of the level, and instead it's focused on trying to find a way to get farther in what it considers a more reliable way of progressing in the level.

>>14053869

Dying should give a set amount of negative points to make the ai choose a different route.

>>14053869

No, the real problem with MarI/O is that the faggot tried to make it learn situations instead of levels by having it make decisions based on what it currently sees, with specific blocks being white or black being tied to randomly generated AND/OR/NOT logic which mutates per generation but stays the same for the run, linked into buttons. If collumn 8 row 5 is white AND NOT collumn 8 row 4 is black press right, for instance. That'd mean that if the Mario's location is always collumn 7 row 4, it'd run right if there's floor in front of it and not an enemy.

Except because of this, it learns to make a single set of instructions that is good for that particular level. It doesn't prodce anything that learns over time to become good at Mario, no matter what you throw at it. Just that one level. It also can't decide "oh this is where I failed, let me try something else when I get there". It just has to get lucky and generate instruction logic that always jumps if there's a wall in front of Mario. And with randomness, that can take infinitely long.

Different situations require different logic. MarI/O is thrash.

>beats stairs

>gets stuck on flag

>reset

Top strats.

>>14054621

Which by pure luck, this genome finally brute forced the stairs once only to not understand to jump for the flagpole.

Now thanks to the way this is programmed, we need to wait until the values for climbing the stairs become dominant in another week.

A guy can teach an AI how to drive a car in a week of training, but this AI can't even get past a flight of stairs after a week and a half?

>>14054647

The AI basically just throws shit at a wall multiple times until it sorta works and advances. The ones that last longer/get further get hybridized to try what they did, but slightly different.

It's either going to get very lucky, or it's going to take months.

>>14054660

I know, which was my complaint about self-teaching AI is often times just a single brute force solution to a problem rather than actually processing something. The problem with the scoring is that instead of reacting like a player would, the AI ends up as a pseudo speedrunner.

It's like science really, you have to implement one variable at a time and in the case of this instead of making a fit all solution the AI "learns" the solution to beating the level. Very little if any of these "situational" algorithms of the neural network would allow the AI to beat 1-3 within a try or two. I think that's why MarI/O did so well with Mario Kart since it was just moving forward, back, left and right while avoiding obstacles and walls.

>>14054660

It should only be mutating after a certain amount of fitness, how the fuck does it not take this into account is beyond me. It's a neural network doing random shit to mindlessly increase fitness, regardless of what's in the game.

>>14054763

I think what makes this one painfully slow is the fact that the "learning" takes so damn long in comparison to other AI. It doesn't actually "learn" anything until after the end of the species, which is now bumped up from 100 to 150. Either way, it's still brute force and it should be mutating itself every 25-50 runs for maximum variation.

Even though this gen might have the fixed stairs issue, the problem is it doesn't know how to navigate the moving platforms. This is also partly because most AI systems are too basic to handle moving objects and spatial memory.

>>14054794

> It doesn't actually "learn" anything until after the end of the species

That explains why it keeps dying even though it has already gotten to the flagpole, it's still on the first species.

>>14054690

>my complaint about self-teaching AI is often times just a single brute force solution to a problem rather than actually processing something.

You can actually get good results that way. Notably, AlphaGo Zero was trained exclusively on self-play without real match history to rely on and still kicked its world champion predecessor's ass with a fraction of its hardware or training time (4 Google TPUs for inference). A simple change of the given rules and it also became a chess master. MarI/O is a simple genetic algorithm that manages to advance in the game using only the distance travelled in the level as a fitness function, which is impressive in itself or at least was when it was the state of the art. I'm sure there are better approaches, they just aren't being applied to this domain at the moment.

There was also a paper recently that demonstrated image denoising and other image enhancement techniques, without any kind of training dataset, by only tweaking the input layer of a completely randomly weighted convolutional neural network against an input image, then cutting off training once it's producing output that's close enough but not exactly the same as the input. Certain features can be implicitly inferred and generalized to a surprising degree.

>>14054817

Yes, except those parameters are fairly clear - win or lose with what score. Part of what I think stalled MarI/O for so long is partially that there is no real weighting to how well the AI did other than the fitness score, which makes the AI aggressive and highly situational. It would be better to have a system where it knows when a fail happened at a stage and weight that compared to a pass, rather than a raw score count. Another thing that helps the reinforcement learning is the AI playing against itself isn't time critical like actions in a videogame are, since even though the AI did something earlier that worked earlier due to timing might not work again such as the moving platforms.

>>14048507

Math Grad Student here.

My gut reaction is that, given the massive number of iterations, the AI might overfit the data - it might learn how to do "the first level" really well, but it might crap out badly given an arbitrary level.

I'd like to see MarI/O tested on custom-built levels once it is trained, otherwise I would assume that MarI/O "memorized how to beat each level" instead of "learned how to play a mario game."

>>14054621

Oh, someone already said this, and for a different reason.

This is a problem not just because of what you said, but also for any general AI.

This isn't really "learning" anything. It's all timing.

What is going on is that the computer is figuring out "if I press the jump button at 2.436 seconds, I survive a little longer." It's NOT figuring out "if I jump over that little brown dude, I survive." The difference in the first case is that the computer will "learn" to jump over one enemy, but not jump over any other enemies. To contrast, even a two-year-old human child or a dog can "learn" that the same thing has the same results very quickly. It only takes getting killed by one fucking goomba for a kid to realize that you should jump over all of them; but this "AI" has to learn to jump over every single one.

It's essentially like playing Mario blindfolded and with earplugs in, and the only feedback is a tap on the shoulder when you die. All you can do is memorize a series of button presses that lead to not dying by exhaustive process of elimination; you aren't learning anything.

>>14055058

that's exactly what it is doing, it's not learning how to play the game

>>14055569

Once it beats 1-2 it will be revealed that a journalist was playing the entire time.

>>14055569

So this is what hell is like.

>>14055611

Not sure how exactly this works. It's based on Sethbling's one for Mario World, most likely, which could read solid blocks and enemies, but couldn't differentiate between types of blocks and types of enemies. So assuming this is the case, yes, it could probably learn, but I doubt much will carry over from level to level.

>>14055611

Can someone edit that pic so the coil is the 8chan infinity?

>>14055611

it likely won't carry anything over considering it won't even jump at the flagpole

>>14048507

>>14055611

>>14055085

Meanwhile in competent A.I. land

>A.I. understands the importance of not dying

>Understands importance of killing enemies and getting power-ups

>If it hits a wall it will try to jump and failing that it will backtrack

>Able to figure out glitches and use them to its advantage

>Takes what it learns into future levels

>Understands importance of actually reaching a win condition (i.e. reaching the end of the level) instead of just going as far right for as long as possible.

>Can play other games as well

>>14055611

I mean, I'm not an AI creator, but it seems to me that programming an AI to beat early Mario games should be fairly simple (ignoring the warp puzzles in the last Bowser castle) if you have AI experience. Especially if you give the computer infinite lives.

You give the AI a list of goals. Chief among them is reaching the flag, which always lies to the right of the level, so the AI gets a "point" for each frame Mario is moving to the right. The AI also gets a point for picking up a coin or hitting a block, a handful of points for picking up a mushroom or fire flower, and a few more points still for picking up a star. That way Mario will hit blocks as he travels to the right, but never double back for one. He will double back for power-ups, however, which is good as it should make it easier to make it to the end of the level.

You also give the AI some amount of memory for each level. The AI would record each action it takes for each frame of each round up until death. The next round the AI will perform a different action on the second-to-last frame and see if that allows Mario to keep progressing to the right. If it does, the new set of data is kept; if not, the process is repeated a frame further back until some action is found that allows Mario to progress further right. And then you keep all of these outcomes in a heap or tree, and if it comes to a point where for whatever reason progress is impossible on the current path you just regress to an earlier point and start trying actions at that frame instead.

That way you wouldn't even need the AI to know what each button does. All the AI needs to know is if a button action allows it to progress. The AI won't know that A lets Mario jump, what the AI will see is that A prevents Mario from dying in this specific circumstance. That's all it needs.

Of course, that's an extremely, extremely simplistic AI, and I can't speak for how this particular AI is programmed or its complexity. But I think it answers your question in a roundabout way, which is to say…

tl;dr No, the AI will not necessarily be able to take "lessons" from level 2 and transfer them to level 3. It's possible, but in order to tell you for sure I'd need to see how the AI is programmed.

>>14055611

No. It restarts and you get another few days of guessing. This "ai" is absolutely retarded and whoever designed it sucks.

>>14055660

Playfun is great and all but it can't use an actual console because it uses time travel. It's entirely emulator bound.

So if it only learns when it reaches a new species, why is the genome cap at 300?

>>14050871

Because it's easier to do it :^)

Shit is pretty retarded, it isn't learning shit, it's just bruteforcing the levels. The "AI" has no info on the level whatsoever but how far it has progressed. Can't see walls, enemies, anything.

>>14055700

It doesn't learn. It does the equivalent of glorified guess work. It might as well be trying in sequence every single possible key combination until it gets to the end.

When you've seen what the Google DeepMind AI can do over a few generations you'll realize this guy didn't do anything of note and he's just doing this for recognition.

The AI is no more complex than something you'd write for an indie game, it cannot learn.

>>14055716

So its a shit ai?

>>14055699

nigga I ain't saying playfun isn't good it's just an entirely different implementation.

Seth (as far as I remember) made his AI work live, and therefore it works on the fly with no prediction.

>service jones makes another thread

>>14055720

This is like the Jeb! ai that serves no purpose while the superior Tay ai exists.

>>14055724

oh really, smarty pants? MarI/O needs to teach itself every single level, while PlayFun actually learns how to play the game from a single human demo input. What MarI/O is doing is bruteforcing key press combinations and timings PER LEVEL until it nails one that finishes THAT LEVEL. When it has to beat a different level it starts from scratch. It's fucking trash. These kind of AI are good for specialized tasks, but not tasks that require animal like intelligence.

>>14055777

I'm not defending fucking MarI/O I'm saying that playfun can see into the future and determine what it should do next, while MarI/O can't see into the future and thus doesn't have this advantage.

How hard is it for it to encode to try jumping and moving forward at the same time when it gets stuck?

>>14055815

but goyim that would ruin it's learning experience

>>14055660

Oh look, this guy managed to fix the problem that the MarI/0 fag can't, and its only a few extra lines of text for his ai.

This is oddly easy to fall asleep to.

Here's another guy doing it sped up starting at the 1-2 checkpoint. I really think the way the AI is written it will never get past the stairs, hot take I know.

https://youtu.be/330gWOZfOYg

>>14056024

>simplescreenrecorder

>not ffmpeg

step up ur linux-fu, lad

>>14056012

Sad thing is you really want to see the damn thing succeed

>>14056172

it just werks my dude.

>>14055611

>its skills that it acquired in previous levels are not being applied to new levels

Worse, it "forgets" old skills when it acquires new ones. It learned that in order to get over obstacles, it had to jump. But then it came to one of those pink triangles at a corner, where you run up the side of the wall if you keep running, where it had to learn to not jump. When it came to another solid block, it had "forgotten" to jump, and just kept running.

As I said, this is because it doesn't see or understand anything. It just does random shit until it finds a pattern that works, and repeats those patterns until it doesn't work anymore, rinse and repeat.

>>14055679

That would be solved by having an AI imagine future states of the game from what it currently knows. In a sense the AI not only has to learn how to play it, but learn how to self-emulate the game in its imagination by deducing environment rules and physics.

The human mind would be doing this to learn games, because even when you're not playing it, your mind would be imagining scenarios subconsciously to solve problems you had. That's why sometimes taking a break and trying again later is more effective (allows the mind to think freely about the problem without distraction). AIs would need this ability too.

<still stuck at 1-2

probably isn't given actual sprites to remember

I didn't read anyone's posts yet but op, too fun to watch, ill read thread in a bit just wanted to bump and say they need to update the ai, seem to forget what to do and it seems to have no real intention of getting coins, it should also be able to learn what you can do, like it seems its main choice to get past that low part at the start kinda where the turtles come out from he gets hit every time to make it under instead of sliding or breaking blocks or to try to get above everything where you can just run across

From watching for a bit this ai is dumb, I can tell that its main objective is to make it to the end. I would of made it so it actually test what it can and can't do instead of telling what is bad and ok, ill just tell the ai when you die that is bad and when you can't continue moving forward that is bad and to try to assest the situation to try to establish why what it did caused the bad things to happen then not let it happen again. I'd make it a full blown ai to base everything off of shape and movement as well. Also some other anon says to not restart the level, I think that would get very boring but I do think the ai might learn faster but this idea is based off the fact the ai views restarting over again is a bad thing…which i have no clue wether it does or doesnt… I bookmarked that shit OP, thanks.

>>14053869

Wasn't this problem why Google switched from their gamesolving AI optimizing for in-game points to a win/loss ratio instead?

>>14056762

I didn't read that far, that was what I was kinda talking about, thanks for pointing it out

>>14052037

>dumb

>>14052886

>Still counts as a view to YouTube

You really think someone would do that? Just go on the internet and tell lies?

>>14056172

>>14056198

kek

>loop of a handful of pants-on-head mario run videos

>artificial ai

y'all got trolled

Update:World 1-2 has been beaten twice on the Long Plays Land stream.

1-3 is next.

>>14056843

well I guess ill find out in a couple days when I check back in on it to see if it got any smarter, and someone above proved if you can't down it it's live then…

>>14056870

>Save state at the pipes

>Couldn't even beat the whole level in 1 go

What a joke

>>14056900

It would if given enough time. The original stream makes to the last block. The other one just seems to be sped up with more save states.

Besides, that level was already beaten by NES AI, with one change that stopped counting downward falling as progress.

>>14055660

>>A.I. understands the importance of not dying

I was shocked, but the realized it doesn't mean the importance of living.

Well Mario seems to be doing better today than yesterday by just a tiny bit, doesn't die as often at gap be for entering pipe. Wait wait wait, I found a flaw. It does a save state at pipes but when it comes out that final pipe be for the stairs and then gets stuck…. Shouldn't there be a fucking save state right at the fucking stairs then?WTF? Who ever made this bot must of lived like when nes first came out or he is retarded as fuck. You guys see the fucked up chat box in the stream? Lots of trolls and no one is blocking them or anything

once the genome that gets stuck on the last step of stairs becomes the majority, maybe it'll finally evolve and clear the damn stage. hell, there are instances where that happens in a row now unlike a few days ago where it'll kill itself before the pipe over and over.

it's hard to drop this shit when the AI is clearly learning something albeit at a snail's pace.

beep boop are you winning son?

>>14064103

48 65 6C 6F 2C 20 77 6F 77 6F 72 6C 64 7E 36 24

30 39 30 39 31 36 35 30 30 30 31 39 35 32 33 0D

0A 3A 31 30 31 45 35 30 30 30 39 30 39 33 36 35

30 30 38 30 39 33 36 34 30 30 32 46 35 46 35 46

54 68 69 73 20 69 73 20 69 73 20 61 20 68 65 78

63 69 6D 61 6C 20 74 75 74 6F 72 69 61 6C 21 46

didnt somebody make a better mario ai years ago?

>>14056012

>is written

Murder yourself, brainlet

>>14053584

>implying I don't want eugenics for humans

>>14068863

What else would you prefer him say, you unlovable chucklefuck?

>>14069218

probably, choded. like coded and chode, because that anon is a chode

>>14053174

The problem is that they are running it in real time. It should be accelerated 1000x or more in order to take advantage of the AI's capabilities.

I don't like that each one plays from the beginning of the level even it dies.

If you don't make them keep up the skills from the previous level then they're just going to develop level-specific solutions for each one, each level will take 6 days, and they won't ever find a general solution that can play the entire game.

>>14069636

Guess that means it's learning though

My issue with these kinds of things is that they aren't really learning anything they can apply elsewhere. It knows how to get through the level and has issues with higher level game mechanics.

They are basically game journalists but more patient.

>>14069218

Evolved. The AI isn't written at all, it "evolves" over time using a process similar to natural selection. None of the shit is hardcoded in, and the problem we're seeing in the game isn't that it's too dumb or badly coded, but actually a fundamental problem with evolution itself; it takes a lot of time.

Most of the posters in this thread don't even understand anything about AI. For example >>14055660 doesn't get the fact that that specific AI uses a completely different algorithm (a deep-learning neural network) that is trained using prerecorded play-footage. Furthermore, the "neural network" in that has a completely different structure, the differences which (I think) are explained in the first few minutes of vid related.

TL;DR most of the anons in this thread are brainlets who can't into AI

>>14070044

Also, the AI in MarI/O EVOLVES (as I said), it does NOT "learn". That's the point of the I/O in the name, it means Input and Output. What basically happens is the AI(s) is given a simplified version of the screen. Then, at random, different positive (if) and negative (if not) links are generated and connected to a specific input (block) on the screen and an output (controller button). An example would be a connection saying IF there is an enemy (black block) at block on the screen (x,y) jump.

An AI with this connection is simulated and placed into the game and then assigned a fitness score (how far it moved to the right and how fast it moved to the right) along with a bunch of other randomly generated "AI". Whoever has the lowest fitness score is culled for the next "generation" (just like how an animal with low real-life fitness would die in an incompatible environment) and more random connections are made. Eventually over the generations the connections or the lack of connections are optimized so that the "AI" (which really isn't even an intelligence, it's more like a dynamic sheet of rules that a computer follows) is able to complete the game.

Although I hate CGP grey, he's reddit as fuck, vid related is a pretty good babby's-intro-to-AI.

Why is everyone in the chat a tremendous trap-sucking fag

So I was looking over chat earlier today, not sure if it's confirmed but seems that it glitched through a block in 1-1 and lost all of its species in the process somehow, that's the only explanation I've seen why it's been sitting on 1 species this whole while. Also the fact it may not have even been possible for it to do 1-1 legitimately.

>>14071333

Someone also added to that, that that's why it seems to like to attempt just juggernaut its way through blocks, because of that earlier glitch. Could be bullshit.

>>14053584

>implying hatred of eugenics is not a fabrication of post WWII social conditioning.

>>14071926

You're a skinwalker.

Bamp. Mario still gets stuck on the 1-2 flagpole block, but at least gets to it most of the time

>>14070044

I don't think this Mario AI has enough input/sensors. I think its just using a genetic algorithm to string together random gameplay actions in a set order. Its not learning to play the game, its learning the right combination of inputs to beat each level.

>>14068797

The vids are already in the thread, but its amazing how it can effectively play other games that it wasn't even optimized for. With a little tweeking, I think even the games the playfun ai fails at it could be pretty good at playing.

>>14048507

>for almost 9 days

So you're saying the ai learned to increase its defacto speed?

>>14082060

Watching that other one play Pinball is some sweet science.

do any of the AIs actually take into account visual input? You know, the things that real intelligence uses to learn to play the game? Like if it could recognize patterns on the screen like stairs it would be able to get shit done so much faster.

>>14082188

Depends on the implementation, these AIs just learn by a reward function that may take points if they get hit or die, and get points by completing a level or going further right.

They don't really need to know what enemies are, what's the level layout or how the game works, they just learn when and how long they should press A or what direction should they hold on the digital input.

>>14051958

Calling this thing an "intelligence" is a misnomer.

>>14048507

Your mom's vagina. They both fall in and are never seen again.

>>14082751

Source?

>>14083346

俺のバチュルユーチューバーあいちゃんはこんなに可愛いわけがない

>>14056024

>gay sex with hats on

Thank you for lightening up my bleak day.

>>14056024

Why don't you check the link, my good man.

https://8ch.net/tech/ddg.html

>>14084951

>tfw i didn't catch my retarded boy clear it

how the fuck was that 3215 fitness achieved though?

>>14084951

well done you glorious dumb ai

>>14084971

That's how far it managed to move to the right

>>14084951

But, they forgot to save a recording of this feat.

16 days

Still on 1-2

>>14086351

Right subscript and superscript are equivalent to below and above sigma.

>>14089912

Oh hang on I must stand corrected.

>11 FUCKING DAYS

>>14087346

Anyone new: Original link is done. Please use this new stream link.

>>14087346

either the streamer died or he's just waiting for the AI to get more than 75% clear rate in a generation.

>>14055587

Honestly it would be more impressive to train a journalist to play a game than it would be to train an AI

>>14051958

>Hand it to the jewgle robot that plays chess. It would probably just copy a TAS input file to play through it.

THAT IS NOT HOW ALPHAZERO WORKS YOU STUPID FUCKING RETARD NIGGER CASTRATE YOURSELF

>>14048771

but what if the AI considers the amount of time alive as a success rather than the priority being reaching the goal, then by all means this is a gold strategy in the eyes of the AI.

>>14053101

<~we can do better

I cant wait for retarded sjw activists to scream about how ai is real human with feelings and rights instead of a fucking code immitation

>>14054621

Is that how it works?

Holy shit, that's remarkably fragile. There's no way this shit is getting out of the local optima.

>>14089976

Looks like it is now doing 1-3

>>14099564

Go tell that to /pol/

I dare you