Re: [computer-go] Modern brute force search in go

[Vincent Diepeveen <diep@xxxxxxxxx>]

At 17:31 7-11-2004 -0500, Don Dailey wrote:
>
>> GCP has been doing now a few measurements with R=3 with his simple go
>> program and already has a branching factor of 7.45
>
>Vincent, I  don't see how it's  possible to benefit from  null move in
>GO.  I used  to think this was  be perfect for go, but  I have serious
>doubts after thinking about the problem.

I heard the same fairy tales in 1996 when i asked why Hsu wasn't using
nullmove in deep blue.

Frans Morsch just has proven nullmove to be great as he won the world
champs 1995 thanks to recursive nullmove. Apart from recursive nullmove,
fritz was a major shit compared to other programs at the time.

>In chess, null move measures  threats by essentially doing a pass with
>a depth reduced search.  This gives one side 2 moves in a row and this
>is deadly in chess and effectively measure threats even with seriously
>reduced depths.

>But in GO it seems to me that  even with the advantage of 2 moves in a
>row the reduced  sub-search depth makes you blind.   That's because it

If you search 29 ply, with R=3 at a 19x19:

d4        1
nullmove  2+R
q4        3+R
nullmove  4+2R
d16       5+2R
nullmove  6+3R
q16       7+3R
nullmove  8+4R
k10       9+5R
nullmove  10+6R
q10       11+6R

In short 6 moves you get for free.

How would a professional evaluate this board position?

It's a forced win with black probably if you study hard.

Nullmove works just as good in go as in chess.

Just be sure you have a good evaluation function...

>usually  requires really  long sequences  of specific  moves.   I know
>there are  cases where it would  be quite beneficial  in filtering out
>junk, but I don't believe it would be the rule like in chess, it would
>be only the exception.
>
>Does anyone have  any evidence that null move can  be pretty useful in
>Go?  I ask this question honestly, I don't know the answer.

At this moment nullmove in go just doesn't prune enough to see deep enough.

More drastic methods are needed to get to that 30 ply unforced until
hardware is a billion times faster.

>- Don
>
>
>
>_______________________________________________
>computer-go mailing list
>computer-go@xxxxxxxxxxxxxxxxx
>http://www.computer-go.org/mailman/listinfo/computer-go/
>
>
_______________________________________________
computer-go mailing list
computer-go@xxxxxxxxxxxxxxxxx
http://www.computer-go.org/mailman/listinfo/computer-go/