[Freeciv-Dev] Re: AI (again!)

["Peter Schaefer" <schaefer@xxxxxx>]

>> One way to scrutinize the rules would be playing them with adaptive
>> learning AIs and see if they converge to different strategies.
>
>But if rule is really unfair/unbalanced,
>it nearly allways means
>that human players will understand it,
>but no AI will.
>
>This is nice test for rule unfairness/unbalanceness.
>( AI don't go for huts on poles, for example,
>  nor run for Michelangelo nor Bach )
>
Nonsense. Exactly the opposite will be true.


>
>Idea of ``adaptive learning AIs'' means (in most obvious implementation)
>that you make small random diffs to AI's behaviour, run game, test
>which AIs were the best, and start cycle again.
>This is nice for adjusting heuristical parametres, but how can it
>create any completely new behaviour.
>
You pour oil onto the flames. People will be telling you soon about
"emergent behaviour" a buzzword that describes simple rules
leading to more intelligent looking behaviour.

However, in my cellar room lab experiments
for a simple board game, the AI genepool population
soon consisted of a single strategy;
probably this is mostly due to the small size of the
genepool that you can process in reasonable time.
There would be a better chance to fill the game with a member of
one of 7 different families ( for example a difference in some
parameter(s), but better in code ) every game, and then to look
whether the different families converge in behaviour or can play their
style.