Re: [gnugo-devel] joseki

[Heikki Levanto]

On Sun, Jan 27, 2002 at 12:44:15PM +0100, Arend Bayer wrote:
> However, maybe we should compile a joseki dictionary ourselves from
> free available game collection, in a manner similar to extract_fuseki?

I am not too sure. I think a good joseki dictionary is much more than a list
of commonly played sequences. There should at least be some info on when to
choose this or that variation, and what to do if the opponent deviates from
the sequence.

When to prefer one joseki over another is a tricky question, and would
ideally depend on our knowledge of our own strengths and weaknesses. No
point choosing a fighting joseki, if we can not handle the fight that
follows it.

As long as our engine will not automatically find the optimal answers to
people who deviate from josekis, we must have the most common trick plays
and blunders in the dictionary as well. Otherwise people quickly learn the
points where deviating from the joseki costs a little bit, and induces gnugo
to play a much bigger blunder.

For these reasons I feel that we must at least keep the hand-crafted joseki
base. It may be possible to extend it automatically, but we still need to
add the conditions when to choose what, and the deviation sequences.

Another approach would be to make gnugo learn from its experiences. When
ever it meets a situation that deviates from its joseki database, it records
the move it chose, and the result of the game. If the human opponent plays
consistently the same variation, gnugo's database would grow to recognize
that as a bad move. Unfortunately there is a long way from a joseki mistake
to a lost game, and lots of noise can creep in between... 

Even if the computer learning is too ambitious, we could note the deviating
moves in the game records, extract the most commonly seen ones, and add them
into the dictionaries by hand.

-H




-- 
Heikki Levanto  LSD - Levanto Software Development   <address@hidden>