Genetic Algorithm Deathmatch
Seeing this Slashdot story on this use of machine learning in a computer game, made me wonder if anyone has tried to do anything like use genetic algorithms to control and evolve AI players in something like a Quake deathmatch?
Chuck a bunch together with random algorithms... the players that survive longer get duplicated and modified. The appealing thing about that is that the players would be in a fairly complex environment, which might help drive towards fairly complex behavior, and that it could be fully automated - the players, the selection and modification. The complexity of the environment includes the 3D environment, other AI players, their strageties, how they handle the various types of situations that come up, etc etc.
I'm not that interested in spending the time to seriously look into this but I did this google search, which seems to include a few such things, such as the work described in this paper.
You'd need a better algorithm than just those that survive or you'll breed a bunch of players that just go hide in a corner somewhere away from any action.
ReplyDeleteHi,
ReplyDeleteI haven't thought this through that deeply, but I don't think that's necessarily true. Say that every X minutes a player has survived, an offspring is created for it. If nobody dies then the environment is going to get pretty packed: the hiding spaces will run out and there would be, I think, an advantage going around picking out all the players whose strategies are to hide away in one place. Even if hiding out gave a player an inherent benefit when fighting someone out in the open, there are still more ways of being out in the open then being hidden away, so those hidden away could still end up in the minority.
James.