[Computer-go] I need an off-the-shelf final position live/dead evaluator
ddyer at real-me.net
Sun Nov 28 11:02:19 PST 2010
At 07:16 AM 11/28/2010, Michael Williams wrote:
>On Sun, Nov 28, 2010 at 4:15 AM, Dave Dyer <ddyer at real-me.net> wrote:
>> At 10:39 PM 11/27/2010, David Fotland wrote:
>>>Accurate scoring, even at the end of a game, is very difficult. You have to
>>>read accurately, and evaluate semeai and seki.
>> Yup. I spent years developing the capability to score endgames
>> at the point where humans typically leave them. Getting within
>> a few points of correct 95% of the time is achievable. The other
>> 5% you will either make whopper mistakes or never terminate.
>> This is now pretty old, but I don't know of any more recent or better
>> results. http://www.andromeda.com/people/ddyer/go/scoring-games.html
>No one here or at that link mentions whether they are talking about
>Chinese scoring or Japanese scoring. It seems that the there would be
>a significant difference. For insnance, if I wanted to score a
>Chinese endgame starting with no code, I would probably write an MCTS
>move generator that played the game to the bitter end and then score
>it using simple area counting. But you can't do the same thing with
The referenced page was based on Japanese scoring, and treated the
board position as a semi-static position. One part of the process
involved filling the dame in a controlled way, to allow mandatory
responses as necessary, but that was after it had decided what was
supposed to end up alive.
It's a good question if MCTS would get the right answer in those
5% cases - one characteristic is that they are poised on the brink
of life or death, where there is frequently exactly one sequence
that determines the "correct" outcome. It seems to me that MCTS
would tend to say "it's alive 99% of the time' when the correct
answer is that it's 100% dead.
More information about the Computer-go