วีดีโอ

Now make the ai play deep dip 2

Maan, I would love to see this AI learn deep fear for 3 months

Boring

What a video. Enjoyed every second.

Hi, Love the video. At about 20 mins you talk about rewarding the AI if close to the world record line. I am just wondering if this 'ham strings' from finding more optimal lines once it finds the path.

Would love to see this IRL in those Formula-E cars (or whatever they called) obviously wont be able to wall bash, but would be nuts to see it turn into a competition. Its the ultimate showcase of AI+Human achievement together. Humans do mechanicals and AI does the driving. That really turns it on its head as you would assume it should be other way around with a machine making a machine and human driving it.

At this point you just w Should rent some AI server for Training

Nobody cares bro. It's not AI at all. It's literal basic scripting

The work you are doing is really awesome, never give up on it, RL has always fascinated me. Can't wait to see some human intervention creating supervised IQN!!

Incredible results! Congratulations!

You should make an ai for the f1 games and then sell it to f1 teams to train their riders, easy million

this is awesome. I wonder if there’s a good way to “direct” the AI to take a general pattern (like bouncing off walls or leaping a shortcut) by incentivizing it to stay within a certain distance of an example run (in a satisficing and not a maximizing way.) This could have similar benefits to how human input with machine help via Foldit sometimes has better results than straight AI. Often the best protein folds use human intelligence with machine tools (Shake and Wiggle, AlphaFold.)

Ok the part where the AI stops at the bridge to maximize its reward. This happened over a decade ago with an AI that was trained to play a video game and beat it as quickly as possible. It started freezing or crashing the game in order to "not be punished"

So basically, instead of playing for 2 years, you worked for 2 years on some robot to play for you instead.

Top AI is now solid at chess--almost unbeatable, but at Go AI loses to skilled amateur humans with some anti-computer strategies, and at poker and DotA2 can only play pro level at artificially simplified versions of the games, while playing significantly worse than pros at the full verions of the games, and also loses to the best human pros even at some of the simplified versions. People in the field of AI have a habit of misrepresenting and overstating their achievements. AI has not "mastered" or "conquered" nearly as much as claimed. Not yet, at least.

This remeber me of half size micromouse. Nice work!

Dumb. And truly sad. Watching a computer repeatedly playing a video game is just about the most worthless, wasteful use of one's time, maybe ever. Go outside. Learn a trade. Take cooking lessons. Get in a few fistfights. Go on dates and get your heart broken. Get an actual career where you accomplish something and provide a bebefit that is truly useful to the world--farming, carpentry, etc. If you are fascinated by racing, learn to drive a real car and get some track experience. You know, go do some real man things. Computers are supposed to help make our actual lives better, to improve our practices and procedures, not to replace our experiences with digital serotonin and dopamine hits on little blue screens.

how is this done and how difficult would it be to set it up?

Not expert ai but yes if its trained properly ai will smart enough knowing the timing and limit of something under its control, to give best output

So just to out this in perspective the worlds vest time holders were beaten by a fruitfly brain.

Kind of funny that an intelligence biased towards "immediate rewards" is better at racing ;)

23:13 Not us being destroyed by a FLY

I hope we get to see the guy eat his hat

dudes teaching skynet how to chase us down

I'm guessing your low res "could you even imagine" image is at least half the pixels of a gameboy. We did alright. You even have greyscale, back in my day it was black on neutral grey/brown.

AI or Machine Learning?

19:56 He essentially programmed anxiety

TAS proves how superior humans are. Humanity, FUCK YEAH!

So... you cheated! Using humans to teach the "ai!...

awesome stuff! IS the AI able to be used on any races? Or is it only trained on 1 race at a time and thus 'memorizes' the single best route for a single map?

14:19 I laughed so hard! 🤣

16:35 your AI *is* a tool, so it *is* a tas. sorry to break it to you :c

theres gonna be a time where accusations float around of people submitting speedruns done by AI whoo boy we are getting there aint me

Did bro just use AI to beat a WR in Trackmania

for people saying he didnt train the AI its just stupid cus he infact made a machine do the runs instead of him doing the runs so its AI

I think the reason AI cant find shortcuts is that in order to find them, the time would have to go down for many runs before going up again. since I presume the ai gets punished for its time going up, that would mean shortcuts, at least complex ones will be impossible for the ai to find unless found early on in its training. its for this same reason that animals cant evolve to have extra pairs of limbs or breath fire. because in order to do that, it would have to evolve a bunch of other stuff first, which would lead to the creature getting "punished" by being less efficient and adapted for its environment. evolution is very similar to ai learning if you think about it that way.

Yosh already beat you to it.

wehen you dont have skill^^

Surely v4 (or 3.5) could just be v3 with higher resolution frames? I think human wr's would be slower if we were playing a lower resolution game too, so I'd expect there to be unpredictable benefits to the existing ai if it has more information input. It would also go some way to providing the AI with a way to look ahead, in an analogous way to how humans actually look ahead in game- by literally looking ahead in the track in any given frame of the race. Since training times per race have already come down so much due to your other optimisations, it seems like increasing the processing requirement by increasing the resolution needn't make training take weeks and weeks. Also final (small) recommendation- collab with Code Bullet! He might have some clever ideas having done a lot of ML projects in different games.

When the AI agents achieves the fastest time, can it repeat it consistently? (Percentagewise)

Are you saying we can train fruitflies to race?! :D

Uh, hey dude. Your AI training gonna make some people time travel to get you.

Have you looked into using cloud computing to run your code on? Can be massive help to the training process

Idk about the rest, but in dota 2 open ai unable to beat pro teams(it cant win a single game). But it winning against amateurs very often, about 98% win rate

could you train an AI to play rocket league? it wouldn't be easy at all but i'd love to see it

Have it try deep dip

Hmm, I have some thoughts about this. As I play Trackmania now, I'm coming to realize that this AI model is probably inadequate for several reasons. Consider what human players do when they drive. They don't _just_ drive, they also _learn._ Of course, the AI learns too, but it's learning the wrong things, it's constructing an incorrect gameplay model in its "head". First of all, it should somehow be able to learn certain essential skills, such as breaking, drifting, speedsliding, etc, not to mention the tricks on all other surfaces and with other vehicles too. Next, it needs enough memory to build a "visual map" of the track. Us humans do it by eye, but for the AI, I recommend logging the exact spatial coordinate where it plans to execute a given trick. And only _then_ it could begin to start reasonably learning. Though, you probably still can't do much about tracks where one needs to severely slow down or go down a very specific and odd route. Solving such tasks requires creative thinking.

It is not intelligence because it does not understand anything, it's just a database. It is a neural network.

Can the ai speedslide?

Using random population and darwinism, can find more local minima/maxima