Quoting Martin Mueller <[email protected]>:
The main limitation of Fuego, and I suspect all other MCTS programs,
is in the playouts, not in the search. I wrote up my impressions
and some analysis after the matches vs strong humans in Barcelona
last year.
Yes, I noted that Fuego has some weaknesses Valkyria does not have in
the playouts. And my feeling is that these weaknesses pop up a lot in
games between Valkyria and Fuego. One example is Nakade in seki (also
with false eyes that are true eyes in a seki).
A recent example of this problem from a Little Golem 9x9 game:
http://www.littlegolem.net/jsp/game/game.jsp?gid=1334704
The evaluation was 70% for Valkyria for a long time but about 50% for Fuego.
Note that I am not sure the version of Fuego is the latest.
The version of Valkyria playing on LG is not the same as tested on CGOS.
I have done some things to try to overcome the noise problem of MCTS
and memory limitations. The games on LG is played with up 12-24 hours
using 2 threads (several positions are searched in parallel as games
are played in parallel on LG) on a 4 core machine.
This version of Valkyria is currently undefeated (there have been
losses in the past due to operating mistake and with short times and
the standard search).
I cannot make a strong statement of scalability of this version of
Valkyria. I just see that a few moves into the games the evaluation is
monotonically increasing at a rate dependent of the mistakes of the
opponents. With white (komi is 5.5 favoring black strongly) the
evaluation may drop slowly as long as the opponents play good moves,
but so far Valkyria seems to be able to make the game complex enough
to make strong players make little mistakes that finally add up to
winning the game.
Note that I describe here i purely anecdotal. The number of games
against really strong player is really low. Still my experience is
that many humans are able to play a really strong opening so many
games have been good tests.
Magnus
_______________________________________________
Computer-go mailing list
[email protected]
http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
