AI Learns How To DRIVE

[@ceefour11]

Bro you need more subs, super underrated.

[@Shivam-Cr-7]

He's right

[@ospy_edits]

Yup

[@HungryKittyCat]

Agreed no more talk

[@Baller628]

"Ai learns NUCLEAR LAUNCH CODES" your welcome for the quality idea

[@Punch_Card]

how do you implement that lol

[@pro100lalka4]

I think you should've let it run for hundred or more generations more, I would love to see an AI not only to learn it and complete it, but to MASTER it, and show some crazy stuff, like cutting the perfect angles on the turns, doing max possible speed at every given circumstances, etc. I recommend you next time to try some game where an AI can learn some far from human abilities, and reach them, it would be so interesting to see how far they can go, just add time variable, so they were rewarded not only for not dying, but also for doing something faster than before! But anyway you're super underrated youtuber, I never liked the content from channel with 4k subs so much, keep it up!😊

[@TylerMommsen]

Thanks for the support man! Sure, I'll try do that with future videos then. Any recommendations for some games that'd be interesting to watch an AI do that? Maybe I'll also just try to make a 3D version of this which would be pretty cool.

[@gattasara]

I also believe they should

[@CatSplat9018]

@@TylerMommsen I personally would like to see AI trying Crossy road (loving your content btw, keep up the amazing work)

[@TylerMommsen]

Thanks for the support! Okay I'll keep Crossy road in mind for the future

[@picxelz2284]

@@TylerMommsen You can also check out the trackmania AI videos they show some pretty cool stuff. Not sure if its the exact same as what you're doing but yea.

[@GemedetAdept]

I really appreciate that you leave in and show off the actual programming process in addition to the end result. The result is always fun, but as a programmer I love seeing someone else's process and how they figure out solutions.

[@sketchychillandchill]

Put it in a random track from a pool on each generation

[@mr-nafari7355]

Good job man, you obviously deserve more subs. Keep up the good work 👏

[@minheepark4896]

You should add like a hour long Timelapse of all the generations and stuff in a separate video for people to watch the “evolution”

[@TylerMommsen]

Maybe I'll consider opening a second channel for timelapses only

[@minheepark4896]

@@TylerMommsen ok 👌

[@CuberrOfficial]

Rare footage of me driving

[@ammarbumadan2258]

What an amazing video dude, I'm not a coding nerd and i found this super entertaining to watch. Keep at it 💪🏾

[@TylerMommsen]

Thank you, I appreciate it!

[@The0Stroy]

Next step - they see each other and can collide.

[@ThunderFound]

Hey! Would be great if you record a video explaining how Reinforcement Learning works. You are very underrated

[@TylerMommsen]

Thanks! Yeah ill release something in the future

[@gabrielnieto6813]

Great content!! hey few ideas, now that they can finish the lap, what if you add a new reinforcement of time so that for each generation they learn doing the lap faster is better, it would be interesting to see lap record, faster way to get there etc

[@hinakremec]

Man when you will be a big youtuber, remeber us viewers who were there since the beginning

[@Hellllloagain]

here

[@Poppieandboefje]

this guy is underrated

[@dahonigerdnussbrot6934]

Pleeeeeeez make the AI go faster! I would enjoy this extremely! (Also... u need more subs, how is your channel so small??)

[@TylerMommsen]

Sure no problem! Maybe i'll upload a follow up to this video or another one but in 3D 👀

[@qfurgie]

can you do this again, but instead of distance, it’s time based? they don’t die when they hit the edges, and they are killed off after say 30 seconds, then if any car beats that arbitrary threshold, that begins the new kill point (lets say 25 seconds) until the cars reach a threshold that is physically impossible to beat and you have best optimized route.

[@TylerMommsen]

Yeah maybe that would have been more interesting. I did implement a time related scoring system with regards to each cars average speed but the main focus was distance traveled. Your method probably would have been better

[@SMGSEBASTIANSMG]

Hella underrated bro

[@Psygenz]

Good shit, keep up the amazing work! 👍

[@PhantomGamer80]

There are cars!

[@Almancea]

Edge surf next?

[@TylerMommsen]

I'll definitely consider it for a future video

[@Tiger_Cub0]

“F1 cars don’t drift” Max verstappen in an alpha tauri

[@rkom4104]

You should've tried to add a punishment for each seconds of existence for cars (if this AI training model supports it though). The cars will do their best to receive minimum punishment points so that's equal to drive as fast as possible

[@AgastyaSondhi]

Could you add laptime as a reward so the cars are incentivized to go faster?

[@TylerMommsen]

Yes you can, you can use laptime and the average speed over the duration they are alive

[@Ryan_Swivet]

I love your stuff man

[@timur.shhhhh]

0:21 JavaScript? There is no JavaScript, only Assembly

[@Tantalium1]

I thought this was gonna be trackmania for a minute xd

[@Croissan_]

Bro hearts almost every comment Edit: even this one!

[@TylerMommsen]

Damn right i do

[@eli_0625]

Very Nice!

[@RealPacse]

I'm really interested in AI and was wondering, do you recommend any tutorials for getting started with NEAT in Unity? I have experience with Python, C and HTML, but have no experience with JavaScript or Unity. Any pointers would be appreciated :) Your video's are amazing!

[@JEXC_NITROTYPE]

You should do trackmania.

[@WhatAmIDoingHereOnYoutube]

as an actual bahraini thank you for recognizing my country😭😭

[@krstev29]

Question: If one car makes it, for the end of time, will you get always perfect runs after that?

[@TylerMommsen]

Yes, but only for the one car that makes it as he will be pushed to the next gen. The way NEAT works is it allows worse performing players (cars) to go on to the next generation because maybe they have some genetics that will be more useful later on many generations later, so that's why even though one car is significantly better than the rest, the next generation will not be a ton of copies of that car

[@pika28samson]

Nice vid :D I am also training a similar AI, may I know the fitness data, input, and details on specific training values? Tysm ❤

[@DebugBug]

I have a question, and if someone can explain, that would be great. I once tried to code machine learning, and it technically worked, but at that time I didn't know much about AI, and really, the ai worked by a thread. So here is the question: What operation goes on in the node? Ik most of the time its a weighted variable, but how do you decide? is it predetermined? if so, you have practically made everything. If its not, then how does it decide what mathematical equation to put on that node? thank you to whoever answers!

[@TylerMommsen]

A node will take the sum of all the weighted connections to it as an input and then it will apply what is called an activation function to that total to normalize the input. Common examples are sigmoid and tanh. Then it's output will simply be that value. The decision on what activation function depends on the situation. Sometimes you want values between 0 and 1 and sometimes you want it between -1 and 1

[@DebugBug]

@@TylerMommsen but does the AI have to decide the function itself, or is it predetermined. Anyways, Thank you so much!

[@TylerMommsen]

@@DebugBug No problem! The function is predetermined. At least for all my AI's that is. The activation function i used for my nodes in this particular AI was the tanh function to map values between -1 and 1 which worked perfectly for left/right steering and reverse/accelerating

[@Litger]

Hello Tyler i want to learn about neural networks and teaching ai cool things like you,can you teach us how can we do these or can you sell lessons? (Sorry for bad english)

[@חננאלרועיארבל]

Do you program the NEAT network from scratch? If you do, is there a chance you can make a tutorial on how to make a simple network? If not, can you provide a link to where you learned it?

[@TylerMommsen]

Yes, it's programmed from scratch in JavaScript. In the future I plan to make some tutorials or course on how to write it up completely, but for now you can check out The Coding Train on KZbin, he has a full 64 video playlist on AI which helped me a lot to learn. But before that make sure you are familiar with a coding language

[@kylemorris5338]

@@TylerMommsen How did you manage to wire a JavaScript AI into a Unity project? Most of what I found on a quick google search was a combo of "you shouldn't do that", "you CAN'T do that", and "there was maybe a way that halfway worked but it was discontinued by the company".

[@TylerMommsen]

@@kylemorris5338 oh damn sorry I made a mistake in my previous reply. I already had my NEAT code written in JavaScript from previous projects and I just had it converted into C# for Unity.

[@kylemorris5338]

@@TylerMommsen ah, that makes a lot more sense. Cheers!

[@aceldama]

just curious (and too lazy to read teh whitepapers), how does your neat algorithm choose when and where to add new nodes?

[@TylerMommsen]

NEAT adds new nodes and new connections based on a random mutation chance. For this AI it had a 5% chance of adding a new connection, 1% chance of adding a new node and a 80% chance of changing the existing connections weights. These happen with every new generation. As for where, new nodes will always be placed in the 'hidden' layer which is just between the inputs (left side) and the outputs (right side).

[@bloxfy-bh7rz]

Nice!

[@The_Doorser]

Day 2 of asking him to make snake from chrome

[@erkot000]

But does the ai know how to drive or did it only memorise this one track

[@TylerMommsen]

In general since the AI uses rays as inputs, it will be able to use those inputs on any track they are placed on but in order to make them really good, it would be more ideal to train them on multiple unique tracks

[@erkot000]

@@TylerMommsen Thanks for the answer

[@tusharjadhav721]

nice bro...

[@demongod9120]

Bro can you tell me how I can do such code recently I have been trying to learn machine learning and ai stuff in python but I still can't figure it out and yet I am still interested

[@TylerMommsen]

Sure, firstly make sure you are fully familiar with Python or any other language of your choice and then you can check out The Coding Train on KZbin, he has great videos in JavaScript and some in Python for writing AI. You can also check out Tech With Tim. He primarily focus on Python and he makes great Python tutorials and he has tons of vids on AI/Machine learning stuff. Hope this helps!

[@lucryc]

nice vid

[@egbfilmz5839]

ai learns chess pls

[@random_precision_software]

Have u got the project files?

[@TylerMommsen]

In the description

[@random_precision_software]

Yea I've seen the source code but no project files. i.e grfx​@@TylerMommsen

[@Mineninja12]

why did you not just use rigidbody 2d?

[@TylerMommsen]

I did use rigidbody2d, what do you mean?

[@ChimbroDaPro]

Let's GOO

[@lurelleru]

Молодец 👏👏👏👏👏👏👏👏👏👍👍👍🎉🎉🎉🎉🎉🎉

[@Shivam-Cr-7]

Yep i did like lat video

[@ninacooks1]

👌👏👏👏👏👏👏👏

[@gslive6197]

Please sir make auto play free fire game bot for level up in android

[@bryceblazegamingyt9741]

Everyone watch this video on 0.25 speed on loop while ur out doing whatever to boost the algorithm!!

[@flameaep]

if even ai can, why women can't?

[@cantaloupeduckofdoom]

wowowoowow

[@Tybek1]

Nice copied thumbnail