AI with Funny Pictures

[ark626]

For my studentical work i have crafted a AI trough evolution.
The main system is called Neat or the advanced one is hyper Neat
Currently it is doing a really shitty job on playing connect 4

But thats only cause the opponent is an algorithm called Min Max.
This algorithm checks all doable steps about their chance of winning and takes the best one.

Its pretty complex but i had quite fun creating it.
When im back ill explain how it works in general IF anyone has some interest

The Pictures of the AI's, their generatornetworks and the current best Game AI is checkable at
91.228.53.18

PS: The program is still learning so the pictures are updated each 15 Minutes.

I would love to build that system as mod for a SoG enemy

 

[GarlicJelly]

This reminds me of this guy: http://www.cs.cmu.edu/~tom7/mario/
It's pretty cool to see AI play games since they can sometimes come up with a different way to do thing because they can rely on pixel perfect strategies. I'd love to see a TAZ run of SoG once it's completed.

 

[ark626]

Yeah the point in SoG is, to train AI's sucessfully you need a "Fitness" function that evaluates the Sucess in a number, so the AI knows if an approach is good or bad.

In Mario the Fitnessfunction often just is how fast you get from left to right and how far.
Pretty easy. In SoG you can beat that down to a task, that a AI learns.

This is the reason why i would love to play around in a mod with the enemies, training them and enhance their behavior.
As you said the AI is able to learn pixel perfect moves BUT also has errors, and those errors would be intresting for the player and his playstile

Because the evolution would fix overused errors but also leading to new errors

My system currently is Hyperneat which is basically overkill
Normal neat system should pretty much do the trick, altough i need to refine the input the AI gets.

Input is the own HP, the surrounding (Players, walkable ground, which moves do player make, other enemies, which moves do they.)

This input needs to be refined out of the game, so the ai doesnt need an input neuron for each pixel, because that is pretty much a speed issue.