On 17-nov-08, at 14:42, Michael Williams wrote:
My reasoning is that more deterministic playouts are going to be
stronger playouts (assuming they are done right), and so should
contain less noise. But because you don't want to be playing the
same playout over and over, you need plenty of randomness near the
start of the playout.
I see. It's an interesting idea. I did observe that making the
playouts too deterministic does hurt.
I was thinking that the justification was going to be something along
the lines of: the probability of playing a move should be based on
the estimate of it being the best move. Since early in the game there
are more open possibilities, a certain fixed pattern is less likely
to be absolutely the best in the beginning than towards the end. But
this doesn't hold for all patterns. Making a ponnuki, for example.
It's clearly more likely to be the best move early in the game.
Anyway, thanks for sharing the thought.
Mark
Mark Boon wrote:
On 17-nov-08, at 13:36, Michael Williams wrote:
You'll probably have to test more than one percentage on each
type. It's possible (and likely, I think) that 50% could result
in worse play while something like 20% results in better play.
Also, I'd like to re-submit my idea of increasing that number as
the playout progresses.
Yes, I may have to try more than one likelyhood. Maybe something
like 75%, 50% and 25%.
I have been thinking a bit about your suggestion of making
playouts more random early in the game and less random towards the
end. Why do you think it would help? Is it just a hunch, or do you
have a specific reasoning?
There are some instances I could think of where it might make
sense and others where it might be the opposite. But since I don't
want to complicate things too much at first I think I'll initially
stick to a fixed number.
Given the myriad of possibilities I'm starting to think if it
wouldn't be better to use genetic programming to breed a playout
strategy. I don't know anything about genetic programming so I
have no idea how suitable it is. Somehow it seems to me that if
genetic programming would ever be useful to computer-Go it would
have to be for something fairly simple and contained like MC
playouts. Maybe that's a project for another life though ;-)
Mark
_______________________________________________
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
_______________________________________________
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
_______________________________________________
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
