Hi, Ingo.
Indeed it is a puzzle. As you well know (and you have commented on in
the past) my Amazons program, Invader, displays the same behavior --
including during self-play. Since MCTS, under most implementations, is
truly oblivious to the margin of victory the only conclusion one can
come to is that, for some reason (?!), close finishes end up with
slightly higher win rates. I suspect it has something to do with the
structure of the tree some levels down from the root. For example,
perhaps positions that are "winning big" also allow slightly more
chances for a catastrophic loss as well. But it's hard to make a cogent
argument to that effect.
You got me thinking about this problem as it relates to Wanderer, my
MCTS Breakthrough program. It is not clear what "margin of victory"
would mean in the case of Breakthrough, but one reasonable definition
might be the difference of the distances to goal for each player. It
does not appear that Wanderer prefers closer games, but maybe that's
because all games tend to be very close anyway. The same with Havannah.
I don't think I see many "intentionally" weak moves that keep the game
close. Maybe this issue manifests more conspicuously in games involving
territory?
-Richard
On 05/31/2013 03:27 AM, "Ingo Althöfer" wrote:
Hello,
especially in the early years of Monte-Carlo Go it
was often observed in games between MC(TS)-bots and humans
that bots won by the smallest possible margin, 0.5 points.
We all know that this is not a bug but a feature ;-)
For a long time it was my impression that this phenomenon
was typcial only for bots-vs-humans, but not for
MC-bots vs. MC-bots. But now experiments with other games
make me believe that wins by small margins happen often also
for MC-bots against each other.
Who has experiences or explanations for this (in Go)?
Ingo.
_______________________________________________
Computer-go mailing list
[email protected]
http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
_______________________________________________
Computer-go mailing list
[email protected]
http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
