Adding a prior through a 3rd player should alter your results. There's
minimal data on how A compares to B, so since the data for dummy says
they're equal, you should expect results closer to one.
You could do priors that simply set the initial guesses. That won't
alter the results. You also could increase the number of games between
A and B. That'll push the results back to theoretical.
Are you sure the connect pattern is still valuable once the other
patterns are taken into account? It may be that without other
knowledge, it works well, but that a full pattern set can better
differentiate when a connection is good or not?
Sent from my iPhone
On Jul 24, 2009, at 9:33 AM, Isaac Deutsch <i...@gmx.ch> wrote:
Hello,
I have a question regarding the paper by Rémi Coulom. There is a lin
k to it here:
http://computer-go.org/pipermail/computer-go/2007-December/012557.html
My first question also relates to that mailing list post. Rémi says:
"My suggestion would be to test your code with very small amounts of
artificial data. For instance, start by two players A and B, and,
say A beats B twice and B beats A once. Gammas should converge to
gamma_A = 2 * gamma_B."
I tried this test and it converges to the right solution when I have
no prior. When I have a prior (which means 2 competitions where each
feature wins and loses once versus a dummy feature), it does not
converge to the right solution. Instead, gamma_A approx. = 1.66 *
gamma_B. Is this normal? Also, are there other artificial tests I
could try?
Secondly, I used the technique explained in the paper to calculate
ELO data for various features, including 3x3 patterns. I get good
convergence for all features except the 3x3 patterns. I have made a
chart showing this:
http://files.getdropbox.com/u/1298345/elo.png
It is most notable for the "connection pattern" (an example 3x3
pattern that connects 2 groups), but it also applies to the hane and
empty triangle pattern. It seems like the gamma value isn't totally
wrong, but it doesn't converge to a stable value but to zero
instead. Any ideas how I can fix this, or where I should start
debugging?
I used about 4000 pro games and every 4th move in these games for
training data, so about 100k trained positions. I do merge
symmetrical patterns.
Regards,
Isaac_______________________________________________
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/
