Petr Baudis wrote:
I wonder now, do you use separate set of gammas for simulations and node
biasing? Since I've found that the performance seems very bad if I don't
include some time-expensive features, since the gammas are then very
off; I will probably simply generate two gamma sets, but perhaps it's
enough to do some trick like merging features by computing weighted
(geometric?) averages?
I learn two sets of gammas separately for the two sets of features.
You say you consider patterns that appear at least 625 times, does that
mean that they are _played_ at least 625 times, or that they are among
the move candidates at least 625 times? I'm using similar filter based
on the former, but judging by the numbers it is likely you do the latter?
Yes, number of times they are candidate counts.
For your 9x9 tests, were you using the gammas you harvested from 19x19
games? I'm finding that especially border distance gammas seem very
unsuitable for 9x9.
Yes.
P.S.: It's fascinating that you've managed so great results with plain
UCT, I've always thought in the past that you are using the patterns
with UCT+RAVE. It's strange so few independent implementations of your
method exist... (I now wish I'd use the endless hours I've spent tuning
my rule-based heavy playouts to rather implement the general pattern
matcher.)
I use AMAF as a criterion for progressive widening.
Given the number of programs that are successful with RAVE, I believe it
is likely that RAVE is a better approach than progressive widening. I
started progressive widening before RAVE was known, so my program is
highly tuned for the progressive-widening approach. I rapidly tried RAVE
and it did not improve strength, but that is probably because I did not
try hard enough.
Rémi
_______________________________________________
computer-go mailing list
[email protected]
http://www.computer-go.org/mailman/listinfo/computer-go/
