Re: An AI program that doesn't learn

[Patricia Hughes and David Elsdon <babel17@xxxxxxxxxx>]

David Fotland wrote:

> At 12:16 PM 6/29/99 +0200, Patricia Hughes and David Elsdon wrote:
> >"P.J.Leonard" wrote:
> >
> >I do think it is necessary to keep track of a group's identity. My program
> has
> >no notion of the history of the game. It determines its move from the current
> >position and nothing more. And I am convinced that this is the right way
> to do
> >it.
>
> I used to think this way also.  But you have very limited time, so you can
> only
> analyze a few cantidate moves.  If you ignore the last move, your program will
> tenuki too often, just because it never looks at local responses to the
> last move.

My program is clearly a different "shape" to Many Faces in that it does
not
analyse a few candidate moves, I always consider the whole board, albeit
superficially. I woud agree that it is probably a useful heuristic to
check out 
local moves first, but I need to think about this. My concern
is that it is also a heuristic used by weaker players and one that
actually
weakens their play. When I play Go I am aware of non-local big moves and
I try not
to get drawn into local responses when the profit there is less than
that of the
big move elsewhere. That is ideally what I would like to code into my
program, but
at the moment I am not able to do that because my program simply doesn't
work that
way - yet.


> My evaluation function knows nothing of the game history, but my full board
> move
> selector is aware of the last two moves, and makes sure that the obvious local
> responses are examined first.  This way if there is a big local response (by
> the full board lookahead and evaluation), it will find it rather than making
> silly-looking tenukis.

Well, I will continue with my idealistic whole board approach for the
time being,
and I will handle the "silly-looking" tenuki problem by making sure that
my
program only makes "brilliant" tenukis.

Regards

David