I have been spending my recent time on trying to make a Neural Network brain as the AI of the boxers in Concussion Boxing.
It's not easy to do and it's very error prone.
Every small change in the neuron network algorithm that is not exactly what is needed can make the whole thing fail to give results.
I even read an article from some Intel guy saying that NN are not suitable for games. I believe it is suitable because there were(very few) games who did this in the past.
I am hoping that once I figure this out I will write an article on my devblog (PompiDev.net) about how to make NN AI work for a real game and specifically a fighting game.
For now let me show you the first step I took in this quest for an NN controlled boxer.
In this video the red boxer is controlled by me and the blue boxer is controlled by a single neuron.
You will see some counter running down on the top left of the screen. This is how many frames are left before the learning phase is over.
The numbers bellow it are weights and you will see they slowly change to positive numbers which is the correct solution for this specific neuron.
The only purpose of this neuron is to maintain a constant distance between the player and itself.
What do you think? Would an article about how to implement NN in a real game will interest you?