Hi, I wanted some feedback/suggestions for a really cool idea I had that I'm trying to program right now.

I'm making an A.I. for Toribash using Neural Nets. That means that the A.I. needs to be "trained" by playing games of Toribash, and getting rated on its performance.. I need suggestions on what to rate it on though.

A neural net works by getting input values (for example, how much damage the player has received), and providing output values (in this case, which joints to extend or contract).

The problem is, neural nets are too stupid to learn how to do cool stuff right away, you have to teach them simple things at first and build on them in tiny tiny steps. This particular implementation I am using (called NEAT) is very powerful and can plug into any sort of problem with a little work.
---------------------------------------
Ok so my big questions:

Does anyone know what information I can pull from Toribash using LUA to feed into this neural net program? The neural net needs some inputs and the more useful they are, the faster a good A.I. will evolve.

Should I be training the NN by running individual games of Toribash and using LUA to transmit the moves?

I know Toribash uses a replay system to record all the input for the 2 players.
Has anyone figured out how the replays are encoded? (maybe I could be generating replays with my neural net instead of playing individual matches of TB)

If I have to run individual games of Toribash, does anyone know if theres ways to speed it up or turn the graphics off altogether, since I just want to record some value like damage or distance?



Thanks for reading!


-------------------------------

A program I am making requires me to run many thousands of matches in Toribash in order to retrieve the values of some variables (things like joint positions and damage inflicted on both players, at round end).

Is there any way to speed up this process so Toribash runs as fast as possible? I don't want to watch the matches taking place, I just want as many results as fast as possible.