On Jan 17, 2009, at 5:41 PM, Sylvain Gelly wrote:
For the first difference you mention, as far as I remember it makes a small but significant difference and is one of the main reason I was talking about "tricky details".
OK, I ran a test and after 1,000 games with 2K semi-light playouts I get a winning percentage of 50.6% for your methods vs. mine. Of course it's possible I made some mistake, but my first impression is it makes no difference which way you do this particular detail.
Your ChooseNode is also quite different from mine, mostly because I also still have a UCT component in there. I'll give your method a go one day, just to see if it changes anything.
I've come to understand what you mean by "tricky details", sometimes I see a big difference in playing strength that I find hard to explain given the change(s) I made. Conversely I've been in quite a few cases where I thought something would make a difference, only to find out it all didn't matter one bit.
It's also possible that some deficiencies that would be apparent in one implementation, get compensated for in another.
Some examples: David Fotland wrote he does light playouts with just a few patterns but no tactics. I find that using a moderate amount of tactics actually is the biggest contributor to playing strength (save one or more stones if can't be caught in ladder). However, augmenting patterns with tactical information I found doesn't help at all, even when disregarding the performance cost. Maybe David uses some patterns to compensate for part of the tactics and relies on the faster playouts to compensate for poorer playouts. I'm guessing here, but otherwise I can't imagine why he would forego what otherwise seems to be a big gain in strength.
I also tried to use ownership maps to modify the RAVE value. Remi Coulom wrote in a paper he used ownership information of up to 63 playouts. When I tried something similar it always makes play weaker. Maybe I should use it in a different way, but I haven't stumbled on the solution yet. When I think of it, AMAF information is already something very similar to ownership information. So maybe combining the two doesn't make much sense.
Lastly, in an earlier UCT bot that I made I gained a lot by initially reducing the number of moves and slowly expanding it. After using AMAF it turns out this method hardly gains anything at all anymore.
So the devil is not only in the details, it's also in the combination of the details.
Mark _______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
