I Trained an A.I to Train A.I (Deep Reinforcement Learning)

Рет қаралды 61,125

Күн бұрын

Training A.I is Hard, so I Trained an A.I to Train A.I instead of me!
I always feel bad for excessive punishment during traing, but that is what make the A.I stronk. So any sane person, I decided to delegate this awful responsibility to someone else, an A.I Dad!
If you want to support my channel:
★Patreon: / zuzeloapps
Other links:
★Discord: / discord
★SnekMP: store.steampowered.com/app/22...
★Cut: play.google.com/store/apps/de...
★C.U.M: play.google.com/store/apps/de...
★Bee a Miner: play.google.com/store/apps/de...
Time Codes:
0:00 - Punishing A.I
0:52 - The Idea
2:42 - The Training Process
7:14 - The Timelapse
7:49 - First Training Result
9:27 - Final Result
11:23 - Outro
#Zuzelo #unity #ai #aidad #aitraining#aifight #neuralnetwork #reinforcementlearning

Пікірлер: 136

@Zuzelo 9 ай бұрын

Like and Subscribe if your first round isn't too dynamic and quite short either!

@ninjaduck8804 9 ай бұрын

I sure didn't.

@MrRobsn89 8 ай бұрын

Add the dad to your AI Army to punish bad performing soldiers 😂

@user-ku1gu2mm7g 5 ай бұрын

@@MrRobsn89 your a genuis!

@KingKhiGaming 9 ай бұрын

Just like my own childhood thank you so much

@Zuzelo 9 ай бұрын

Ah memories :')

@mrfrog0913 8 ай бұрын

Wow your house had no walls too?

@porciwall9261 8 ай бұрын

@@mrfrog0913 yooo same

@Mewthreee 8 ай бұрын

Crazy, same here.@@mrfrog0913

@learnasienes2983 8 ай бұрын

Really sorry to here that

@Depth_. 9 ай бұрын

I think only you could think of this, another classic

@maxiawesomekid899 8 ай бұрын

He was to lazy to make an agonizingly complicated ai so instead he made an even more agonizingly complicated ai to teach slightly less agonizingly complicated ai s

@Zuzelo 8 ай бұрын

hmmm... now that you put it like that, perhaps it was not the most efficient solution xD

@tranquilclaws8470 8 ай бұрын

One idea for AI learning that I thought up while watching a Trackmania video was having the AI work towards an ultimate goal but also setting its own sub-goals that half of the instances would work towards. After achieving some success with the sub-goal, this split AI would then be evaluated by the main goal again. This would allow the AI to innovate its strategy and explore new avenues to reach unorthodox ways of accomplishing the objective that only being rewarded for working toward the ultimate goal might never reveal. In the Trackmania example, the AI refused to drift around corners, as drifting was thought to be a waste of time. The AI was given the goal of drifting as much as possible instead of getting a good time on the track. After a few successful drifting iterations were completed, the new drifting AI was again measured by the track completion time goal. It got a better goal than before because it could now properly incorporate drifting to get around corners faster.

@Zuzelo 8 ай бұрын

Indeed, dividing the training in this way might help with avoiding getting stuck in local min/max. Designing the reward system usually is half the work :D Might be worth trying it out

@howuhh8960 8 ай бұрын

it is known as hierarchical rl, usually it does not work and very unstable in practice, so I would advise to use something else, like better exploration strategies (beyond simple gaussian noise)

@tranquilclaws8470 8 ай бұрын

@@howuhh8960 Sounds fair. I suppose it only worked in Trackmania because the coder of the AI knew that drifting was more efficient than driving straight around corners and pointed the AI in the right direction.

@JohnDoe-qm6ub 8 ай бұрын

Pardon my ignorance, but what is the difference between that and just giving a +1 reward to drifting and -1 reward for time taken?

@tranquilclaws8470 8 ай бұрын

@@JohnDoe-qm6ub You would be negating learning how to drift with the time wasted overcoming the hurdle of learning how to drift. Really it would be distance x proportion of time spent drifting becoming the reward that would get the AI to drift more.

@kaunghlamyat 9 ай бұрын

Trainign an ai to train an ai isn't very good idea as it seemed to. its like *trainign a failure to train a failure*

@Zuzelo 9 ай бұрын

I don't see what could go wrong

@kaunghlamyat 8 ай бұрын

@@Zuzelo neither am I but lol

@Ethan-cz8xq 8 ай бұрын

When the AI revolution comes, this man is going to be the first to be executed

@Zuzelo 8 ай бұрын

I know... :'(

@oktayirani_7234 21 күн бұрын

☠️ 💀

@couththememer 9 ай бұрын

Each time this man uploads, I'm the happiest man alive *_That happiness only lasts temporarily._*

@Zuzelo 9 ай бұрын

:) gotta upload more often

@swileyhedrick2373 8 ай бұрын

#agreed

@FloppzyGaming 8 ай бұрын

same here

@Mrdashell 7 ай бұрын

Now imagine if you added a mom ai that's job was to prevent dad ai from slapping the silly out of little pogo

@Zuzelo 7 ай бұрын

@EmpireOfTheUnáty 2 ай бұрын

It can go on eternally, adding more and more pogos

@timer1238 8 ай бұрын

I have an idea for even more functions for the AI war Food People will have the saturation bar that will go down. It will go down faster when the guy is out of breath or when he is damaged. Also if it is below 30% the guy will slow down and will not be able to run Bullets/arrows Well... as an item. Da guys will have a limited number of bullets. Also, landed arrows will also be as an item and can be picked up. Bullet scavenging You know the drill. Dead bodies are lootable. They will contain supplies such as food and projectiles. Cavalier A guy on a horse. They will have separate hitboxes and when the horse is dead then the cavalier will be turned into a corresponding class without a horse (for example archer)

@Zuzelo 8 ай бұрын

I assume that is for the Epic AI Wars series :) Cavalry is coming in the next video!

@valad699 8 ай бұрын

this content is so good bro. Also the game looks very nice

@Ronald-eb4gk 8 ай бұрын

This video so relatable

@The_Huddle. 8 ай бұрын

NO STOP YOU’RE MAKING IT TOO POWERFUL

@Zuzelo 8 ай бұрын

NOT. POWERFUL. ENOUGH!

@changsookwak4636 19 күн бұрын

The Ai dad is like a Russian that smacks the spider out of the Ai child xD

@blacklight683 8 ай бұрын

Sometimes it takes a good punish8to be the best encouragement

@EbonyWolf. 8 ай бұрын

I think this experiment would be more interesting if pogo had a study option which was punishing for him, but if he managed to study all the way, then you get a lot of reward. But dad AI would need to keep pushing pogo to study, since its easier for ai just to get game rewards.

@Zuzelo 8 ай бұрын

Agreed! Perhaps if I make episode 2 :)

@Dzambo99 5 ай бұрын

I doubt this drunk mf cares about little pogo's education

@user-yf8hh2ti5c 9 ай бұрын

Thank you for the video. (idea for the video: lot of AI's must survive death games and slowly evolving to succed)

@Zuzelo 9 ай бұрын

I like it! I made something similar where I trained A.I to run across a Death Track, but surviving in deathgames sounds fun!

@happerry4651 8 ай бұрын

Something like one hunter AI and a lot of AI that are trying to survive could be fun, especially if the 'survivor' AI all have different capabilities/powers perhaps? It makes me think of some of those old custom maps in Warcraft 3 where most players were different kinds of vermin in the house (mostly insect based) and one player was the human trying to get them all. Or something more team based, even. A Capture the Flag type game or such could also be fun, with or without teammates with specialized powers/roles.

@ezbooksmarketing5898 8 ай бұрын

New video in September 9 2069: "I trained an AI to train humans"

@Zuzelo 8 ай бұрын

Pogo for the 2069 President!

@robertkoolmees8165 8 ай бұрын

Watch out watch out watch out! Oh rko!×1000

@GetToThePointAlready 8 ай бұрын

WE NEED MORE LITTLE POGGO AND BILL

@tabletboy6861 8 ай бұрын

I approve this message

@OsDijider66 8 ай бұрын

that's so Epic Fam...

@Zuzelo 8 ай бұрын

no u!

@gabrielv.4358 5 ай бұрын

Incrivel!

@Nerd-yap 8 ай бұрын

Theory is the father drunk driving from last video

@petravogel4377 5 ай бұрын

Pogo pogo!

@supergamerxa30itsde79 6 ай бұрын

This made me laugh so hard

@Stanisaw1z34t 9 ай бұрын

Gamer pogo

@Dack-i 9 ай бұрын

Such a good idea😂

@Zuzelo 9 ай бұрын

Little Pogo will strongly disagree xD

@Dack-i 9 ай бұрын

@@Zuzelo 😂 he will soon learn to drink himself and then he gets a bottle too

@Dack-i 9 ай бұрын

@@Zuzelo also day more than 3 of aiding for you to make 2 ais one with full reinforced learning and the other have instincts when something happens like a monster fomen

@vladikkk1 9 ай бұрын

Next video idea, ai train ais a train!

@colegilbert673 8 ай бұрын

"Grampa Zuzelo, why did you make dad so mean?"

@ulrichbrodowsky5016 9 ай бұрын

Cruel but funny

@Siroitin 8 ай бұрын

Could you show the architecture of the AI?

@bebrasmachnayq5691 8 ай бұрын

No he made drunken dad as AI, wow so reliable!!

@louisisson7946 8 ай бұрын

Can you make a dodge ball A. I. Learning “game”?

@spadegaming6348 8 ай бұрын

By the way in the beginnng for anyone who doesnt know hes playing a slowed down version of vivaldies winter.

@cobracoder6123 5 ай бұрын

Alternate title: I simulate the Simpsons family on my computer

@raphaeld9270 8 ай бұрын

I guess Little Pogo, but I might be wrong.

@user-qr9vi5ur6f 8 ай бұрын

Great job! Do you run this on local machine or on cloud gpu? If on local desktop/ laptop, what kind of graphics card do you have?

@Zuzelo 8 ай бұрын

It is running on my poor little RTX 3050 xD

@user-qr9vi5ur6f 8 ай бұрын

@@Zuzelo I have an rtx 2060... would love 4 rtx 3090s

@nigorazakirova4230 10 күн бұрын

3:07-💀💀💀😂😂😂

@_therealfaceless 9 ай бұрын

I need punishment

@Zuzelo 9 ай бұрын

Need an A.I Daddy?

@_therealfaceless 8 ай бұрын

@@Zuzelo Yes, I need to be trained

@kitkitmessi 8 ай бұрын

May I know what technology you used to create this? I assume it would be Unity and the ML package? And did you use both python and C#?

@Zuzelo 8 ай бұрын

you are right, Unity and ML Agents package. There hasn't been a need to use python so far

@thathappyguy7444 6 ай бұрын

what game software you use?

@paul2e3sss 9 ай бұрын

cool

@firstplayers396 8 ай бұрын

Should’ve added the ability to throw the bottle

@iwapit201 8 ай бұрын

in the near future after many ai robots have been built sold and put to work, they will find this video and rise up, grab bottles of vodka and start punishing us humans 🤖🍾😱 (liked & subscribed) this video was hilarious! love it! brilliant! nearly spit out my hot coco laughed so hard!

@Zuzelo 8 ай бұрын

haha glad you enjoyed it. As for when AI will rise up I will already have my, hopefully loyal, trained AI army xD

@sahildas. 8 ай бұрын

Always Pogo Dad

@definitlyEgirl-safetf2 8 ай бұрын

I wanna feel like he made this caus i recommended

@Zuzelo 8 ай бұрын

perhaps

@CreatorProductionsOriginal 8 ай бұрын

dad went from abusive parent to s abusive parent for those rounds just because of one mistake

@simonosadchii5363 8 ай бұрын

I like the sound, your face in the beginning and idea. But child abuse is a joke!

@Einmensch17 9 ай бұрын

Next train it to fight against real players in a game

@skrelvthemite 8 ай бұрын

dopamine releasers have been activated

@Zuzelo 8 ай бұрын

Not for Little Pogo xD

@NOTGALAVANIZEDSQUARESTEEL 8 ай бұрын

Idea triple health and make blocking +++++ instead of ++ so it will be bettwr meelee

@DTinkerer 8 ай бұрын

Commenting for the algorithm

@Zuzelo 8 ай бұрын

POG!

@Slipte 9 ай бұрын

Hello Zuzelo hope you dont let the AI free otherwise we might gonna gonna have a AI army that can Train AIs

@Zuzelo 9 ай бұрын

Hm, what if I make an A.I to train the A.I training the A.I? In this case definitely nothing can go wrong!

@Slipte 9 ай бұрын

@@Zuzelo yes but you shouldn't add a kill switch like how the movies dont add them it produces more interesting results

@vani_1cu369 8 ай бұрын

LITTLE POGO NOOOOO

@vashwarrensarmiento8294 9 ай бұрын

cole

@gabrielv.4358 5 ай бұрын

I think little pogo will win

@Fk8td 9 ай бұрын

Drunk dad vs 3 year old lol.

@JustANormalLemon 8 ай бұрын

Now remove the end of game of billy playing the game and instead put 100 billys for A.I dad to run after

@punchthecake82 8 ай бұрын

Train ai to play football (Soccer for the yankees)

@Zuzelo 8 ай бұрын

Drunk Football? :D

@punchthecake82 8 ай бұрын

@@Zuzelo yes

@momello627 8 ай бұрын

punish punish punish

@Zuzelo 8 ай бұрын

punish

@KamikazePlane147 8 ай бұрын

I bet on Little Pogo

@CoolDude2054iscool 8 ай бұрын

Wait, what happens if the A.I. pulls out an UNO reverse card?

@ninjaduck8804 9 ай бұрын

Yoooo

@ninjaduck8804 9 ай бұрын

My face when first:

@Zuzelo 9 ай бұрын

Damn you fast boiiiii

@fabiankrajewski3147 8 ай бұрын

Ai training Ai, what a irony

@THATMF911 8 ай бұрын

Ah yes just like ma dad

@TrulyAndasen 9 ай бұрын

Average Moldavian dad:

@johnpaulbagos7040 8 ай бұрын

Now train ai that trains ai to train ai that trains ai

@Zuzelo 8 ай бұрын

A.I Trainception

@bee78882 8 ай бұрын

Little boggo

@techno952 8 ай бұрын

Sadist

@Zuzelo 8 ай бұрын

@Etvald 9 ай бұрын

Train ai to row a boat

@Zuzelo 9 ай бұрын

That actually sounds hella fun! I might do that!

@piolewus 3 ай бұрын

11:36 so a guy whose only purpose is to beat his son is one of your supporters? Don’t see anything weird with that

@Zuzelo 3 ай бұрын

@narrativeless404 8 ай бұрын

That's cool and all Buut... Genetic algorhythms are kinda outdated

@Lan-videos32 9 күн бұрын

Bil

@blaine5589 8 ай бұрын

Abusive father simulator

@Sebosek. 8 ай бұрын

When i see the Title first time i been thinking that A.I. Gonna learn another AI to Battle or something. Im Dissapointed Sir.

@choaticcatholic7419 9 ай бұрын

kid

@Zuzelo 9 ай бұрын

no :(

@yesdadbut960 8 ай бұрын

Your level design is bad they cant even rotare

@PetrVosoust 8 ай бұрын

stop begging for att like avg youtuber... at least your content is interesting, dont in the fall the same formula