[computer-go] Computer chess in past to Go software in future

[Vincent Diepeveen <diep@xxxxxxxxx>]

At 14:47 26-10-2004 +0100, chrilly wrote:
>>
>>If I understand this correctly, you can stay within specs and have a max.
>2x
>>speedup..
>>
>I do not know of any official XiLinx spec/documentation on this topic.
>Generally hardware runs faster when cooled down.  I have seen a short video
>were someone cools down the CPU with N and then it runs for 30 secs at
>double frequency. During my physics studies we had also our N-fun in the
>lab. But we have a practical system in mind which plays 24h a day chess.
>Actually for bigger cluster just normal air-condition becomes a problem
>(especially under Abu-Dhabi-conditions). I do also not know of any practical
>FPGA-Cluster which is cooled with N.
>But good old Crays had sometimes special cooling. Also the program Dark
>Thought ran at the World-Championship 97 in Paris on a specially cooled
>Alpha-machine (was not N). As far as I remember at 660 MHz instead of

633Mhz production chips (which were not yet sold)
which kryoteched at -40C or -45C ran at 767Mhz

>500MHz. Quite an effort for about 25 Elo points.
>The fastest Intel-Chips ran at that time at 230 MHz.

Most played at a 300Mhz PII just like i did.

AMD k6 ran at 233Mhz. AMD basically delivered 200Mhz chips and later tried
to upgrade to 233Mhz chips at mainboards which just went to 200Mhz.

In 1997 it definitely mattered to get another few ply. 

Diep searched 8 ply at the time.

Compared to deep blue's 10-12 ply which we find pathetic now, 8 ply was of
course not having a chance.

Branching factor of most programs was about a factor of 10 in 1997.

1997 is the first year that at the Aegon tournament many programmers
gathered and the general consensus was there that piece square tables
didn't work anymore for chessprograms. Especially a French programmer who
was searching 10 ply with piece square tables (Virtual Chess) clearly
indicated that above that depth he didn't come further with that type of
knowledge.

Go software still has a long way to go, to come in similar situations.

Not only because there is more possibilities in go, also because all the
software is too hybrid in search. 

There is no need for a local quick tactical search simply. It's far better
to have 1 program with 1 evaluation function which very selectively extends
certain branches to see whether groups are hung.

So effort has to be done to make that search quick with the knowledge it
has and do big effort to write go-knowledge to limit the amount of moves
tried in the quiescencesearch. You can for example evaluate a group with
limited liberty as near hung forcing the side with 1 liberty to play moves
to escape in qsearch, which gives the desired working of qsearch.

1997 was the year that in the chessworld most learned that aspect.

In 1999 many previously preprocessor programs were indeed non-preprocessor.

What happens in the go software there now is very comparable to the old
1Mhz single chip programs, for example from Eric van Rietpaap. Who searched
n-3 ply with normal search and last 3 ply with a quick tactical search.

>Chrilly
___________________________________________
>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/