AI learns to Beat any Game #gamedev #unity #unrealengine

Показать описание

Deep reinforcement learning is a technique used to teach an AI a new skill. We've seen this used a lot recently to beat any game and make super strong player but it got me thinking... Why are game developer not actually using them for their game ?

Happy to see you here rare description reader. If you saw this write "the apple does not fall far from the tree".

Valem

Рекомендации по теме

Комментарии

I think also. If the AI is learning, right? While a normal player is like a multiple choice thing so the AI gets confused because the player is doing something they're not use to and they begin to break down from confusion

FareezGamingV-pijo

When i went to college a decade ago, reinforcement learning was beginning to get popularized (better and cheaper computers meant you could actually run those simulations).

I was an avid fan of rome total war so i wanted to teach AI how to command soldiers and just run military tatics.

I ran for over 6 months and hundred of thousands of interactions. The rewards were the hardest part to figure out. But at the end, it worked.

After that i made a little game where you had 500 soldiers and fought 500 AI soldiers.

AI won everytime single time. The game got popular in my course and every day i had dozens of people trying to beat my AI, i even added a pause button to give people time to think but no matter what they did, AI responsiveness was imediate. You couldn't flank or anything.

I ended up putting trees and a field of vision. For the first interactions you could win by ambushing using the trees. Then AI learned that and always sent a few scouts to make sure no enemies were hiding there.

My problem became how to make them dumb. I then resorted to have 3 difficulties based on how many interactions i've made. The more AI trained, the better they got. So i used few interactions as a mean to have a lower difficulty. But it got tricky too because i had to play a match to know if AI was dumber enough lol, there is no way to know without playing, there is no code to debug.

It was really fun. I probably could port it to C++ and SLD3 (i used C and SLD at the time). Felt like a god having my little soldiers beat everyone. I even added a cheer animation because they were really good soldiers.

thesilentgamer

Yes, but deep reinforcement learning is very good for nonplayer characters if you know what you're doing. The trick is to make sure the non player character has certain restrictions that keep it from being unbeatable

erobusblack

the apple does not fall far from the tree

imagineyournamehere

I want to make a Gorilla Tag ghost that learns by watching other players and talking to them then...

I'LL INFECT IT AND MAKE IT SCARE CHILDREN!! Jk 😂

Vilzio