I was doing an in-depth review of Hunter's paper over the extended
weekend... I think I now see the value of the MM algorithms, and the
understand the divergence problem.
MM has the nice property that convergence is guaranteed. Newton-Raphson
(N-R) carries no such guarantee. The paper includes
On Dec 21, 2007 8:53 AM, Rémi Coulom [EMAIL PROTECTED] wrote:
Hi,
Minorization-maximization is a simple optimization method, and I agree
that it is likely that more efficient algorithms can be applied.
Newton's method implies estimating the inverse of the Hessian matrix.
Really computing
On Dec 21, 2007 8:53 AM, Rémi Coulom [EMAIL PROTECTED] wrote:
Hi,
Minorization-maximization is a simple optimization method, and I agree
that it is likely that more efficient algorithms can be applied.
Newton's method implies estimating the inverse of the Hessian matrix.
Really computing
On Dec 21, 2007 10:03 AM, Álvaro Begué [EMAIL PROTECTED] wrote:
I am sure MM is a perfectly good algorithm for this purpose, but it has
the serious down side that I don't understand it. :) I do understand the
general idea behind it and how it works in some simple cases, but I don't
know
Jason House wrote:
Given that doing one parameter at a time may be less ideal, I don't
know if my method would really inherit those properties or not.
Probably not, because the Hessian has significant non-diagonal values.
But I expect it would still converge in less iterations than MM.
On Dec 5, 2007 4:44 AM, Lars [EMAIL PROTECTED] wrote:
I have some questions concernig this paper of Remi:
http://remi.coulom.free.fr/Amsterdam2007/MMGoPatterns.pdf
@Remi: How many iterations you had used?
Anyone of you have similar or other experiences with the algorithm?
I seem to have
I was trying to come up with my own algorithm to maximize likelihood and I
am having a hard time getting it all in my mind. I managed to write a
working algorithm for the case of logistic regression, but it was kind of
brittle and I didn't know how to extend it to the softmax case, which is
what I
On Dec 20, 2007 11:43 AM, Jason House [EMAIL PROTECTED] wrote:
I seem to have more time to think than to code lately. I believe I've
derived an alternate update method.
Thinking more, I realize I messed up a three things...
For one, Newton-Raphson requires
new gamma - gamma = -*L/**L
On Dec 20, 2007 5:39 PM, Álvaro Begué [EMAIL PROTECTED] wrote:
I was trying to come up with my own algorithm to maximize likelihood and I
am having a hard time getting it all in my mind. I managed to write a
working algorithm for the case of logistic regression, but it was kind of
brittle and
On Dec 20, 2007 10:36 PM, Jason House [EMAIL PROTECTED] wrote:
On Dec 20, 2007 5:39 PM, Álvaro Begué [EMAIL PROTECTED] wrote:
Over lunch I thought of another way of doing it that would be very
general and easy to implement. Basically, I can compute the log-likelihood
for a particular
Jason House wrote:
On Dec 6, 2007 11:38 AM, Rémi Coulom [EMAIL PROTECTED]
mailto:[EMAIL PROTECTED] wrote:
Jason House wrote:
This may serve as a good test of if there is enough data to assign
values to the patterns.
I did not mention this in my paper, but you can
On Dec 6, 2007 11:38 AM, Rémi Coulom [EMAIL PROTECTED] wrote:
Jason House wrote:
This may serve as a good test of if there is enough data to assign
values to the patterns.
I did not mention this in my paper, but you can rather easily estimate
uncertainty margins around Elo values. This
On Dec 12, 2007 2:59 PM, Rémi Coulom [EMAIL PROTECTED] wrote:
Do you mean a plot of the prediction rate with only the
gamma of interest varying?
No the prediction rate, but the probability of the training data. More
precisely, the logarithm of that probability.
I still don't know what
On Dec 12, 2007 3:05 PM, Jason House [EMAIL PROTECTED] wrote:
On Dec 12, 2007 2:59 PM, Rémi Coulom [EMAIL PROTECTED] wrote:
Do you mean a plot of the prediction rate with only the
gamma of interest varying?
No the prediction rate, but the probability of the training data. More
On Dec 12, 2007 3:09 PM, Álvaro Begué [EMAIL PROTECTED] wrote:
On Dec 12, 2007 3:05 PM, Jason House [EMAIL PROTECTED] wrote:
On Dec 12, 2007 2:59 PM, Rémi Coulom [EMAIL PROTECTED] wrote:
Do you mean a plot of the prediction rate with only the
gamma of interest varying?
On Dec 12, 2007 3:31 PM, Jason House [EMAIL PROTECTED] wrote:
On Dec 12, 2007 3:09 PM, Álvaro Begué [EMAIL PROTECTED] wrote:
On Dec 12, 2007 3:05 PM, Jason House [EMAIL PROTECTED]
wrote:
On Dec 12, 2007 2:59 PM, Rémi Coulom [EMAIL PROTECTED]
wrote:
Do you mean a
Jason House wrote:
On Dec 12, 2007 3:09 PM, Álvaro Begué [EMAIL PROTECTED]
mailto:[EMAIL PROTECTED] wrote:
On Dec 12, 2007 3:05 PM, Jason House [EMAIL PROTECTED]
mailto:[EMAIL PROTECTED] wrote:
On Dec 12, 2007 2:59 PM, Rémi Coulom
[EMAIL PROTECTED]
On Dec 12, 2007 4:27 PM, Álvaro Begué [EMAIL PROTECTED] wrote:
Clearly I'm missing something, because I still don't understand. Let's
take a simple example of a move is on the 3rd line and has a gamma value of
1.75. What is the equation or sequence of discrete values that I can
take the
Lars wrote:
Anyone of you have similar or other experiences with the algorithm?
I use at runtime the same Bradley-Terry formulas Remí introduces
in his paper. That is a huge advance compared to naif urgency
scores because it gives a measure of how hard it was to win
for a move candidate. But I
Thanks for the file! This should be very helpful when I try to reproduce
results.
It looks like you are not taking advantage of symmetries. For instance,
88|0|17.033168
88|1|12.263955
and
164|0|17.388714
164|1|25.862695
Are identical except for swapping the roles of white and black (88 ==
Oh, I see you have applied the symmetries, but not the swapping of roles.
Still, this should probably be done and cut the number of gamma values in
half.
Álvaro.
On Dec 6, 2007 7:13 AM, Álvaro Begué [EMAIL PROTECTED] wrote:
Thanks for the file! This should be very helpful when I try to
But then you lose information on player-to-move, right?
On Dec 6, 2007 7:18 AM, Álvaro Begué [EMAIL PROTECTED] wrote:
Oh, I see you have applied the symmetries, but not the swapping of roles.
Still, this should probably be done and cut the number of gamma values in
half.
Álvaro.
On Dec
Oh, I didn't notice at first that the player-to-move was encoded
seperatly from the pattern shape.
On Dec 6, 2007 9:53 AM, Álvaro Begué [EMAIL PROTECTED] wrote:
On Dec 6, 2007 9:31 AM, Chris Fant [EMAIL PROTECTED] wrote:
But then you lose information on player-to-move, right?
No. What I
On Dec 6, 2007 7:13 AM, Álvaro Begué [EMAIL PROTECTED] wrote:
88|0|17.033168
88|1|12.263955
and
164|0|17.388714
164|1|25.862695
Are identical except for swapping the roles of white and black
Curiously, the gamma values in your example are way different
17.033168 vs 25.862595
and
On Dec 6, 2007 10:13 AM, Álvaro Begué [EMAIL PROTECTED] wrote:
On Dec 6, 2007 10:06 AM, Jason House [EMAIL PROTECTED] wrote:
On Dec 6, 2007 7:13 AM, Álvaro Begué [EMAIL PROTECTED] wrote:
88|0|17.033168
88|1|12.263955
and
164|0|17.388714
164|1|25.862695
Are
Yes! You are write. I haven't mentioned this. It's a good idea to swap
them all in the form that black has the move-right. Thank you!
I'll fix that. But not today ;)
If you find any inconsistencies in the data please let me know!
By the way: I forgot to cut the file at the end. You should only
Lars wrote:
I have some questions concernig this paper of Remi:
http://remi.coulom.free.fr/Amsterdam2007/MMGoPatterns.pdf
1. Which sense make the prior (Section 3.3 in the paper and where is the
application?
I understand it the way that you put 2 more competitions to each pattern
in the
On Dec 5, 2007 4:44 AM, Lars [EMAIL PROTECTED] wrote:
2. I had run the algorithm on 400 games (including handicap-games) from
the same game-records source Remi used (Section 3.2), but i used an
other month. I concidered only 3x3 shape-patterns and simple non-shape
pattern including
I have some questions concernig this paper of Remi:
http://remi.coulom.free.fr/Amsterdam2007/MMGoPatterns.pdf
1. Which sense make the prior (Section 3.3 in the paper and where is the
application?
I understand it the way that you put 2 more competitions to each pattern
in the
29 matches
Mail list logo
