In all my test with Valkyria it is absolutely clear that strength
improves with more computational resources. On 19x19 it is even more so.
But I do think most programs has a wall in front of the mountain. And
this is *bugs*. Consider what would happen if a line of code in a
program resigns the game at move 17 no matter what the position looks
like 10% of the time. In this case asymptotic scaling will only reach
a 90% winrate against any program (unless the opponent program resigns
often before move 17).
My definition of a Dan level go player was always a go player who has
mastered all basic tactical and strategic skills. In other words,
dan-players are bugfree. There are no positions where they play like
they were clueless, unless the reading required is too high.
MCTS programs are full with bugs of omission. And this is what is interesting.
We need to do two things: First, get rid of all critical game losing
bugs that cripples search. (Programs may still scale in such
positions but extremely inefficient).
Second, playouts must become stronger. Just adding more knowledge to
the tree part of search does not improve global evaluation. It only
improves global move ordering (this is super important for 19x19
though).
I see two ways to go. Sofar peoply has added static knowledge as code
(Valkyria) or patterns learned offline (many of the strongest
programs) or a mix of both.
More interesting is if one can learn local patterns for the playouts
online. Preferrably during normal search. AMAF does this to some
extent but in a very crude and obscure manner. What if one analyzed
each playout carefully to identify moves that should have been played
and use that in future playouts?
Or maybe one could have a really heavy playout that do proper local
search early in the playout and cache the results for future playouts
to speed up things
Or one does goal oriented tactical search on the groups on the board,
before searching globally. Solving all tactical problems such at what
moves are forcing moves, which move can be ignored etc. And then reuse
this knowledge in the playouts.
I do not care if there is a wall or not but I just know that too
little has been done and written about how to make good strong playouts.
To me it is pretty simple. If the playouts make few local mistakes the
problem of multiple local tactical problems on 19x19 will be reduced
dramatically.
But you cannot do that with static patterns, there must be some online
learning that allows the playouts to play critical moves and prune
losing moves close to 100% of the time.
Magnus
_______________________________________________
computer-go mailing list
[email protected]
http://www.computer-go.org/mailman/listinfo/computer-go/
