I also tried dynamic komi with Valkyria a long time ago. It failed. I
did not waste much time on it. Anyway here are my opinions and
intuitions about it.
As usual I am open to been proved being wrong with some empirical
evidence along with a nice algorithm that I can steal and add to
Valkyria. :-)
As already mentioned, one needs to set the dynamic komi to some kind
of magic value, which IMO requires so much insight about the position,
it means that the program should *know* the best way of playing anyway
without the help of the dynamic komi.
Personally I am absolutely happy with how Valkyria plays in handicap
games. It gives large handicaps against weaker players on KGS and
wins. The only problem is to avoid it resigning early because the
situation is hopeless. With handicap on small boards it is so strong
that the problem just do not seem to exist anymore.
Sure it plays ugly moves in handicap games, but this is not different
from even games where it often plays a little bit too "creative" to my
taste.
The problem with MCTS is that they are weak in evaluation. I am pretty
sure that this fixation on dynamic Komi is confounded with the fact
that most programs has quite light playouts, and programs with heavy
playouts still has big holes in the knowledge that leads to delusional
evaluations. I guess the reason that people think dynamic komi is
important is that these bad evaluations can always in principle be
repaired in *hindsight*.
But repairing something in hindsight is useless. If I fiddle with komi
until the program plays a move I like it is not the program that
selects the move anymore. As the programmer I cannot add a "human
expert" to the code.
This is most certainly true for 19x19.
The correct solution to bad plays is to make the program *stronger*.
I am open to opponent modeling such as make the playouts of black in
handicap games weaker. But in this case I think real gain if any would
come from making the statistics more sensitive to the qualitative
difference in available moves, rather than actually modeling the
opponent, by bringing the win rates closer to 50%. Although I think it
would be really hard to degrade the black moves in the playouts in a
realistic way.
-Magnus
_______________________________________________
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
