Abstract:
Monte Carlo Tree Search of the game of Go tends to produce unstable and
unreasonable results when used in situations of extreme advantage or
disadvantage, due to poor move selection because of low signal-to-noise
ratio; notably, this occurs when playing in high handicap games,
burdening the computer with further disadvantage against the strong
human opponent. We explore and compare multiple approaches to mitigate
this problem by artificially evening out the game based on modification
of the final game score by variable amount of points (``dynamic komi'')
before noting the result in the game tree. We also compare performance
of MCTS and traditional tree search in the context of extreme situations
and measure the effect of dynamic komi on actual playing strength of a
state-of-art MCTS Go program. Based on our results, we also conjencture
on resilience of the game search tree to changes in the evaluation
function throughout the search.
It is still few days away from being ready for submission, but I figured
that given the recent surge of interest in dynamic komi at the list, it
might be useful to make an earlier preprint version available:
http://pasky.or.cz/~pasky/go/dynkomi.pdf
Any comments and suggestions are welcome!
--
Petr "Pasky" Baudis
The true meaning of life is to plant a tree under whose shade
you will never sit.
_______________________________________________
Computer-go mailing list
[email protected]
http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
