Re: Digest for lczero@googlegroups.com - 1 update in 1 topic
188 views
Skip to first unread message
Dariouch Babaï
Mar 17, 2023, 10:11:35 PM3/17/23
to lcz...@googlegroups.com
Well I looked (very fast browsing, very sparse), but it appears that
they actually use individual players data with the maia models, and
different general error statistical distribution family (not sure
that is the gist of the increase match to individual play "style"
they claim).
my naive impression for both the average rating target maia models and the more personal one, is that they are using error models, not chess play models to fit.
Anybody gathered how they defined what an error or blunder is. Does maia have both an evolving best play model and an error model, or is there some outside referential deemed best.
Is it training an lc0 with an error models on top? in sequence.. or together..
my apriori skepticism about the model assumptions (if i am not mistaken, which is possible always, me not reading everything, or updating my view by others mean than making hypotheses like here, expeting informative replies in any direction).
is that rating is not really a chess skill characteristic to aim at. I would refer to machine learning results with many suboptimal experts having each their own probably complementary "subpsaces" of expertise... and optimal learning coming from their complementarity as ensemble immitation learning. not exactly as here.. but it brings about the questions that average performance rating may be very mutlidimensional, as much as the game itself... so that learning an individual signature as blunt error model from some putative fixed known best play reference, seems like it is going to blend a lots of information together into noise averaging....
ok that is a stance proposition.. no tomatoes please,... but sound counter points.. would be nice... am i having wrong impression or wrong understanding of the modelling process?
however. compared to maia the claim of the previous post link, is that they improve the "accuracy" of the error model. research is not total fog. news can come out of there.
my naive impression for both the average rating target maia models and the more personal one, is that they are using error models, not chess play models to fit.
Anybody gathered how they defined what an error or blunder is. Does maia have both an evolving best play model and an error model, or is there some outside referential deemed best.
Is it training an lc0 with an error models on top? in sequence.. or together..
my apriori skepticism about the model assumptions (if i am not mistaken, which is possible always, me not reading everything, or updating my view by others mean than making hypotheses like here, expeting informative replies in any direction).
is that rating is not really a chess skill characteristic to aim at. I would refer to machine learning results with many suboptimal experts having each their own probably complementary "subpsaces" of expertise... and optimal learning coming from their complementarity as ensemble immitation learning. not exactly as here.. but it brings about the questions that average performance rating may be very mutlidimensional, as much as the game itself... so that learning an individual signature as blunt error model from some putative fixed known best play reference, seems like it is going to blend a lots of information together into noise averaging....
ok that is a stance proposition.. no tomatoes please,... but sound counter points.. would be nice... am i having wrong impression or wrong understanding of the modelling process?
however. compared to maia the claim of the previous post link, is that they improve the "accuracy" of the error model. research is not total fog. news can come out of there.
Le 17/03/2023 à 08:24,
lcz...@googlegroups.com a écrit :
- Future Chess Program - 1 Update
"brian.p.r...@gmail.com" <brian.p.r...@gmail.com>: Mar 16 07:27AM -0700
Matching moves against a set of players to pick the most likely player to
have made a move is not the same as playing with a particular human
player's style.
Training nets to adopt the style of an individual human player has been
tried several times with Lc0.
The results are "fair" but not "good", IIRC. Moreover, the nets are quite
weak relative to the top strength nets, but that does not matter all that
much in this case.
Accordingly, this area remains a topic of research.
You received this digest because you're subscribed to updates for this group. You can change your settings on the group membership page.
To unsubscribe from this group and stop receiving emails from it send an email to lczero+un...@googlegroups.com.
Reply all
Reply to author
Forward
0 new messages