[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

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



At 23:09 7-11-2004 +0100, Arend Bayer wrote:
>
>
>On Sun, 7 Nov 2004, Vincent Diepeveen wrote:
>
>> >Further, if a human does 20 ply fullboard lookahead, this is backed up
>> >with a lot of local reading that is not included in the 20 plies. The
>> >same will IMO hold true for any strong go program. This reduces the
>> >relevance of this "20" even more.
>> 
>> Forced moves aren't counted in search depth. You get them for free.
>
>What I wrote above has very little to do with forced moves. Most of
>the local reading (life and death, connections, etc.) will never be part
>of the global lookahead (kills usually don't happen in pro games). They
>are needed for evaluation and move selection, "to understand the
>position". And the local reading doesn't consist of forced moves either.
>
>I have the same difficulty as Mark with regards to the claimed branching
>factor of 10.

In 1996 i was crusified at the internet when i said that branching factor
with nullmove would be 3.0 - 3.5 thanks to the increased search depths
(things like nullmove and hashtables start to work real well above 10 ply).

I was a complete idiot according to professor hyatt as Knuth had proven it
to be the square root roughly out of the number of possibilities which
would be 35.

I measured average number of possibilities and it was not 35 but 40.

Branching factor today is <= 2.8 now in 2004.

Branching factor of 10 is for idiotic *todays* go programs. Good ones get 6.
That will go down if more selectivity gets added.

Of course it is a slow change. What you will see is initially go programs
that search like 2 ply and the rest just super selective. And so on.

>Arend


>_______________________________________________
>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/