Michael Williams wrote:
As for the source of applicable positions, that's a bit harder, IMO. My
first thought was to use random positions since you don't want any bias,
but that will probably result in the evaluation of the position being
very near 0.5 much of the time. But I would still try this method since
the lack of bias makes it so attractive. Another option is that you
could take random positions from actual games and generate a search tree
from that position. From that search tree, you could select a few
random nodes to use as test positions. But now we are back to being
coupled to the search parameters -- the same thing we were trying to
avoid by not writing a game engine.
Here are some follow-up ideas I had on how to get meaningful test positions...
The most important positions to evaluate accurately are those near the root of the search tree (because they will be seen first and may direct further
progression of the search. So to generate a "tier-N" test position, take a position from an actual game and make make N random moves. In tuning your
evaluator, tier-N positions are more important to get correct ("get correct" means minimize the error) than tier-(N+1) positions.
_______________________________________________
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
