Petr Baudis wrote:
On Wed, Dec 16, 2009 at 08:57:24PM +0100, Rémi Coulom wrote:
The problem is that I am not using the same set of features for
biasing the tree, and for playouts. Playouts only use fast light
features. The tree part uses slow complex features. In particular, I
use patterns of radius 3 and 4 in the tree, and only radius 3 in the
playouts. When 3x3 patterns are learnt together with r=4 patterns,
they get different gammas.
Interesting, your paper said that you are using patterns up to r=10,
did you find out that anything larger than r=4 is irrelevant in
practice?
At that time (I was not using progressive bias), I found that r=3-5 was
not an improvement over r=3-4. But I re-ran experiments before the UEC
Cup this year, and found than r=3-5 is better than r=3-4. I tried r=3-6,
but the improvement was not clear. So in UEC Cup this year, I was only
using r=3-5.
My feeling is that very large patterns only make a difference in the
opening. After a few moves, they are almost unused, so they make very
little difference in the playing strength.
I have trouble even *nearing* the performance you reported; you say
you can play 13500 games per second, do you have any data on how fast
your engine runs with uniformly random playouts to put this in scale?
My engine does 20k games/s with random playouts, 10k games/s with
random playouts and board implementation incrementally maintaining 3x3
patterns, and 1600 games/s when using probability distribution. There
is some room for optimization, but not to reach 13k games/s...
So I wonder if you have any tips for fast implementation of the
probability distribution based simulations? Do you maintain the
probability distribution itself incrementally over moves, or only
the shape features?
Yes, I maintain a probability distribution incrementally. I also keep
the sum of gammas for every line, and the sum of all gammas. This is
only for non-local moves. So, when I generate a move, I only have to
compute the gammas of local moves.
It is also important to update patterns incrementally. I have a lookup
table that gives gamma for every pattern.
Rémi
_______________________________________________
computer-go mailing list
[email protected]
http://www.computer-go.org/mailman/listinfo/computer-go/
