I looked into this a while ago. First thing I did was research existing work done and found out some guys put together a pretty solid artifical move maker using genetic algorithms and lua scripts for a research project. Pretty neat, but not what we're after.



My vision was to create a program though that would take in a load of replays and learn from those to be able to play a particular mod for example. The first problem with something like that is just amassing enough data to actually train the model in the first place. I put in a request to get a download of all/any aikido games played that were cached or saved anywhere but apparently nothing like that exists.

So we're stuck at the first hurdle there.
Two possible solutions to kick off that first step:
1. Community source the replays. Get as many people as possible to upload all their replays that meet the parameters of what we're looking for. The main issue is that there probably still wouldn't be near enough data and it requires effort for people to contribute. Even if you made a script to crawl the replays section and grab all replays and later filter them for desired parameters you'll probably be short.

2. Create a script that connects to the toribash servers and "watches" games being played in rooms. This involves knowing the protocol to connect to the servers and read game data into a usable format that we can then use to train our machine. I did look into this back a bit and I think with a bit of work should be possible. Then let it run, recording games of the mod you want until you have enough data.



Then from there try to see if we can teach the machine.



This is still not at all trivial. Unlike something like chess where one piece moves to one location each turn, Toribash has 2 players moving simultaneously in 3D space with lots and lots of possible combinations each turn.


For each player:
21 joints with 4 states each = 4^21
And then x2 and x2 for hands gripped/ungripped.
That's...a lot of possible states, and just for one player.



I'm not super well versed in machine learning so I can't really comment on how much of a problem that is but my guess would be that it's on the side of making things harder rather than easier.