Nick Knol wrote:
Hi all,
This is my first post to the list, and I'm pretty new to this, so
sorry if I break from etiquette.
I'm currently working on my senior undergrad thesis project. My idea
is to use Bouzy's dilation algorithm (
http://www.gnu.org/software/gnugo/gnugo_14.html ) to find areas of the
board that form boundaries of influence between different groups and
then concentrate move searches in those areas. After analyzing about
70 dan-level games on KGS I've found a very strong correlation to the
number of moves made in these boundary areas. I define these
boundaries (somewhat naively perhaps) as areas of the board that have
an absolute value of less than 4 after performing an arbitrarily large
number of dilations on the board (>=30 or so).
I've looked around and haven't found any literature on this specific
use of Bouzy's algorithm, so this gives me hope that my idea has some
original merit. Possible applications could include:
-weighting pre-selected moves based on whether they happen to be in a
boundary region
-limiting the branching factor for game tree searches (averages out to
be about 78 over the course of the game, still too many but a big
improvement from 250)
-move ordering for game tree searches
-possible weighting of move distributions for MC-based AI
I was hoping someone more experienced than me could comment on this
and possibly let me know if I'm reinventing the wheel and if any of
these applications seem plausible. Thank you.
-Nick
p.s. if anyone is interested in my data i have some matlab files and
graphs i could post
Hi Nick,
welcome to the list.
Your idea is good, and already well known. I believe Bouzy himself did
something similar to what your suggest (in his paper: History and
territory heuristics for Monte-Carlo Go). I do it in Crazy Stone, too.
You can read in my paper, there:
http://remi.coulom.free.fr/Amsterdam2007/
I use MC territory, not mathematical morphology, but the idea is the
same. On page 8 (table 1) of this paper, you'll see that the "MC Owner"
feature prefers boundary regions.
Rémi
_______________________________________________
computer-go mailing list
[email protected]
http://www.computer-go.org/mailman/listinfo/computer-go/
