Re: computer-go: FPGA

[David Fotland <fotland@xxxxxxxxxxxxxxx>]

I think it is likely that MFGO would play worse if it searched more
moves.  I've seen this kind of behavior.  It's due to an effect that
has not been mentioned yet, bugs in the evaluation function.

Many positions are horribly misevaluated, usually because of group 
strength estimates that are wrong.  Searching deeper could possibly
repair these problems, but searching wider just increases
the probability that the search will stumble upon a position that
it misevaluates in its favor, and move toward that position.

I don't do a general search at the full board.  I try instead to
select only good moves to evaluate, and only search followups to
those move that are relevant to the purpose for making that move.
Then I finish with local and full board quiescence, playing in
unseted groups.  So the move selection at the first ply
is very different from move selection at subsequent ply.

I think Prof Chen said that Handtalk does not benefit much from
full board search, using the same move generator over at each ply.

Tactical search and local life/death search would improve as the
search gets bigger.  I've tested more searching on my life and death
reading, and it gets more problems correect.  This is because the
life/death and tactical evaluations are pretty accurate.  But Many
Faces is already stronger tactically than players with the same
rating.  It needs better full board move choices to really improve.

David

At 03:56 PM 8/30/00 -0700, Mike Gherrity wrote:
>Hi Don,
>
>> Dave, I  am interested in why a  computer would get consistantly worse
>> results   with more computing  power  and   why  you believe it's  not
>> possible to fix.
>
>Dana Nau at U. Maryland showed there were games with this property.  The
>deeper you search, the worse you play.  He called such games "pathological."
>
>It was also shown that scheduling can have such un-intuitive behavior.
Adding
>another cpu can actually increase the time it takes to finish a job.
>
>Although history has shown that more computing power improves play in many
>parlor games, including checkers and chess, it is not clear when this
>assumption may fail.  Perhaps Go has many pathological positions.  Has MFG
>encountered such positions?
>
>
>	mike
>
>