@aiwarehouse

THANK YOU for watching!!:D I know it’s been way too long since the last upload, but we’ve done a lot of work to hopefully start uploading monthly from now on, so there will be a lot more Albert coming, especially humanoid Albert:) As always, if you want to learn more about the AI in this video, keep reading! I’ll do my best to explain everything here, since it’s quite hard to explain it in the video while still making it entertaining.

If you want to join our community of awesome people: https://discord.gg/qDRtuFe5gp

Fan art contest info: The winner of the fanart contest will win a $250 Amazon gift card! The contest ends in 2 weeks on June 15th at 11:59pm EST! The only rule is that it can't be made by AI (ironic, I know), to submit your fanart, post it on twitter using #AIWarehouse. I wanted to say thank you so much to everyone that’s a part of the community, it’s so cool to see how much everyone loves these cubes, and the team and I are really looking forward to seeing all your submissions:) Credit to Echo in our discord server for making the poster!!

Now about the AI, Albert and Kai both had neural networks with 2 hidden layers and 256 hidden units. Most of their observations were in the form of raycasts, we had specific raycasts for each type of observation. Albert for example had a set of raycasts used only to detect Kai’s goal, then a different set of raycasts used only to detect his own goal. It’s standard to use fewer sets of raycasts and just use different tags for the objects they hit (so in this case tagging Albert and Kai’s goals differently then having one raycast set which differentiates based on the tags), but we actually found pretty significantly improved performance when we had a different set for each type (at least in early training). It seems to be easier for the AI to learn, which in our case for this video was very important. Having more sets of raycasts does increase the training times though.

On top of the many many raycasts Albert and Kai had as observations, they were also given helpful direct observations such as the time remaining, their ability availability, distance and direction to the ball (splitting it from a vector into a distance and direction vector helped the AI quite a bit as well), the distance and direction to the goals etc.

For the rewards, like I mentioned in the video we made them very simple, the AI received group rewards when their team scored (even when it was just them on their team), bonus reward based on how fast they score, group punishment when they were scored on, and they also received individual rewards for what we call scoring with style. The faster the ball was moving and the closer it was to the crossbar, the more individual reward they were given.

Now the most interesting thing we tried when we were training them 11v11 was what we called the communicator ability. The agents could turn on and off a collider that only their teammates can see. We actually don’t know how the agents are using this, this was just our best idea for how we could feasibly make an 11v11 game happen, all we know is over the course of their training they were turning on their colliders less and less, I hope that’s an indication that they’re using it sparingly, but admittedly it might just be an indication that it doesn’t help them at all so they learn to just not bother using it, it’s hard to know!

This AI was VERY difficult to train, but it ended up being so interesting. I hope you enjoyed the video as much as we enjoyed making it!!:D

@MJCam1130

3 minutes into the video and Kai is already exploiting a glitch to his advantage. I expected nothing less.

@ibraheemahmed1670

"As a reward, I got you some soccer balls!"

Albert: *dies from joy*

@nii_amart

9:48 Albert's 'bicycle kick' was class!

@CJ0611

11:20 making a barrier from the stadium seats is genius

@karlizkool350yt

I feel like Kai learning about the glitch may have nerfed him. He learned that failing to score on himself lead to wins, promoting aggressive play. When the glitch was fixed, he couldn't learn that defense was the best strat.

@Dementel

Albert: Talent
Kai: Effort

@Lololololololz

9:48 Albert doing a bicycle kick was the highlight of my day

@Luweg01

What i usually see in AI videos like those is that the agents are given super restrictive environments and only learn the super obvious and trivial strategy. This is in my opinion what really makes this Chanel shine above the rest, even tho there are so many great things about it. It is so cool to see them train and improve, coming up with unique strategies and solutions. Also that you are putting them through multiple learning phases makes so much sense and is so cool. Awesome video as usual

@PrickProductions

can't wait in the next 50 years for "AI learns fighting."

@Bill_W_Cipher

Idea for your next video: AI learns to play capture the flag.

@okamiexe1501

Kai breaking a rule 1 at 4:17 is crazy work

@SkippyDev

Times like this is when i REALLY question how this is available for free. The editing is TOPNOTCH!(especially the sponsor segment) and the whole vid is such a cool premise which always got me asking for more! Keep up doin what you do! your work is HIGHLY APPRECIATED TYSM.

(also ALBERT SUPPORTER FOR LIFE 💪❤‍🔥)

@TheDarkScratch

5:47 kai started backshotting albert 😭😭😭

@micahsmith2066

That image at 4:47 with the "my warehouse, my rules" makes this look like a lot more of a horror show.

@balf1111117373

10:20 I like how you can start to see the AI start creating the goalkeeper role. One player is always at the back guarding the goal

@mariosonicfan2010

when i saw that the bleachers could be moved, i was 100% expecting one of them to come up with a strategy of using the bleachers to barricade their goal

@Maxiflame32

You could make a livestream of just 24/7 AIs training against each other like this and have viewers bet which one wins for that day, it'll become an immediate hit

@mickboekhoff

Would love to see a follow-up with different features in the team: taller, faster, wider players & a player who can kick the ball harder. Would be interesting to see if AI adapts a strategy to put the widest or tallest in the goal, fastest in the front to score goals or at the back to get the ball out of the goal & prevent goals, best kicker in front or maybe in midfield for long range goals.

@ibraheemahmed1670

4:15 Kai and Albert are just being friendly