ACG> I have written several computer games using the
ACG> Borland Pascal compiler and
ACG> would like to teach the computer to learn from its own
ACG> mistakes when it plays
ACG> against a human opponent (a la War Games). I am going to try to insert a
ACG> genetic algorithm into the board evaluation subroutine so that the game
ACG> strategies whose parameters produce the most successful games survive to
ACG> the "next generation". Is this a good idea? If not,
ACG> what other method of
ACG> computer learning should I use?
ACG> My first project will be to convert either Checkers, Othello, or Ataxx.
Hmmm. Well, if you do do something like that, I'd suggest you set it up so the computer can play the computer, instead of relying on a human to sit down through a few hundred games until the program gets good. If you have a fast computer and a lot of spare processing time, you could also try a "King of the Mountain" like method... for example:
Method 1 is randomly generated, it starts as king of the mountain. Then another method is generated (Method 2), and 3 games are played between the two. The method that wins the most games stays on, and the other method is erased. Then Method 3 is generated, and it plays the method that won best out of 3 games. Wineer stays on. Etc. Eventually you should end up with something that plays quite well... although it could take quite a long time...
(One more note... in a game like Chess where a draw is possible, a score of 1.5-1.5 could happen in a best of three series, so you might either have them keep playing until the tie is broken, or you could save both methods and have like a "Co-kings of the mountain")
