[Computer-go] 9x9 opening data (was: XMU-FPGA)

valkyria at phmp.se valkyria at phmp.se
Fri Nov 12 01:22:54 PST 2010


Could it be that gnugo plays a little to many good looking but weak  
moves on 9x9 so that it takes you out of any high quality book?

Valkyria has now over 4000 positions in its book, but I think almost  
almost no long lines are in there to beat gnugo. It is simply not  
necessary.

The problem with gnugo is that it often plays lines that are  
territorially inferior with normal endgame, which MC programs will not.

Also the more I add to my book, the more I realize how hard it is to  
get complete coverage.

Best
Magnus

Quoting Fuming Wang <fumingw85 at gmail.com>:

> Darren,
>
> I was able to parse the file without any problem. I considered all board
> configurations in file as GOOD positions, so that any move that can reach
> any board position in the file is considered a good move or a hit. I was
> testing against GNU Go,and wasn't get much hit after 2 moves, which is about
> the same as my small open book build out of 100 professional games, which
> seems strange to me. Anyway, I could have made a mistake here or there in
> rush of time because I was preparing for the competition. I will test it
> again, now I have more time.
>
> Fuming
>
> On Fri, Nov 12, 2010 at 9:01 AM, Darren Cook <darren at dcook.org> wrote:
>
>> > By the way, I tried to use the game record data you posted on the
>> > web, but has pretty low hit-rate during tests. So I had to use my own
>> > calculated one for the competition. It seems strange to me that you
>> > have so many games recorded and has such low hit-rate.
>>
>> Can you tell me how you used it? Did you have any trouble with the format?
>>
>> Perhaps you have some advice for other people starting to look at it?
>> (It is difficult to write documentation as the author, as everything is
>> obvious.)
>>
>> I imagine it is difficult to use it as-is for an opening book, without
>> some processing. (E.g. putting all the positions into an alpha-beta tree
>> and scoring all nodes that way, then storing the joint-best moves from
>> each node.) Did you do anything like that?
>>
>> Poor hit rate could be due to weak opponents who play bad moves and
>> leave the book early? It is also quite focused on the tengen-opening.
>> (E.g. roughly 30,000 games (*) from 5,5; 10,000 from 3,4, 5000 for each
>> of 4,4 and 3,4.)
>>
>> Darren
>>
>> *: where a game is defined as the first 12 moves of the game.
>>
>





More information about the Computer-go mailing list