Actually, this is impossible today, and will be impossible tomorrow.
For an AI to learn by playing, it must have the means to find out the rules, and it must have some kind of evaluating function telling it if the outcome of a move was good or bad.
Now, in cEvo, the rules of the game are not possible to deduce for an AI - the programmer will have to encode them into the AI.
Secondly, the only way an AI can know if it made a good or bad choice is by winning or losing. Any other definition of 'good' will be something the programmer put into it, and will only be a sub-goal, and may be completely wrong the next time Steffen changes the cEvo rules. Now, after a win or loss, how do you decide which move was the crucial one? Is even a single move the crucial one?
I see no possibility at all that an AI could ever *learn* how to win. The possible choices are too many, the exact same situation will never occur again, strategy and tactics are not possible to break down into 'moves' so there's nothing to learn.
Please note that all attempts to have a chess program learn by its mistakes have gone awry, and chess is infintely simpler than a game of cEvo. The only viable chess programs are those that try to compute as many positions as possible, as fast as possible. Add smart alfa-beta pruning to that and you have 'Deep Blue' (the IBM Chess computer).
And chess is a game with full disclosure - both opponents have 100% information of the board. Not so in cEvo.
ai
Posted 2-Aug-08 22:35 by AndersI