Uri Blass wrote:I think that the most candidate for the stall is the fact that LCzero insist to do the same stupid things that alpha zero did instead of trying to improve.
LCzero is weak in seeing mates and they do not consider it as a problem for some reason.
It can miss mate in 1 at least in easy mode.
What about adding a check if there is mate in 1 to LCzero in every position before making playout against itself?
What about simply adding alphabeta to LCzero?
I believe that simple alphabeta when the evaluation at the leaves is simply
expected score at easy mode can do LCzero stronger.
Easy mode is never going to be tactically strong, since it uses the raw policy output to select its moves (i.e. just expands the root node, and simply picks the move with the highest policy prior). It has no lookahead search whatsoever, which is why it's called "easy"
Uri wrote:I think that the most candidate for the stall is the fact that LCzero insist to do the same stupid things that alpha zero did instead of trying to improve.
Sigh. I think you should address your issues with the "stupidity" of A0 to Google.... Maybe they can hire you to help improve its pitiful performance.
About LC0. This is a research project. The aim is to see how far a pure A0-approach will go. As long as LC0 keeps improving (and it still does) I don't see any reason to change strategy.
Naturally everyone can fork LC0 to supplement it with their own ideas. It would be nice to have versions with (1) LC0-search + SF-evaluation and (2) A/B-search + LC0-evaluation(and policy). Daniel's research suggests that the first option would be a disaster. An engine that implements to some extent the second option is Giraffe but Giraffe and LC0 use very different types of NNs.
Ideas=science. Simplification=engineering.
Without ideas there is nothing to simplify.
Uri Blass wrote:I think that the most candidate for the stall is the fact that LCzero insist to do the same stupid things that alpha zero did instead of trying to improve.
LCzero is weak in seeing mates and they do not consider it as a problem for some reason.
It can miss mate in 1 at least in easy mode.
What about adding a check if there is mate in 1 to LCzero in every position before making playout against itself?
What about simply adding alphabeta to LCzero?
I believe that simple alphabeta when the evaluation at the leaves is simply
expected score at easy mode can do LCzero stronger.
Easy mode is never going to be tactically strong, since it uses the raw policy output to select its moves (i.e. just expands the root node, and simply picks the move with the highest policy prior). It has no lookahead search whatsoever, which is why it's called "easy"
I agree with you about the endgame weakness. This is a position from a winning LCZero (ID122) endgame versus Genius, played on the website. Mode was set to hard for LCZero. Mate did not occur until much later on move 112.
[d]5R2/2P5/7B/3K4/3P3k/8/8/8 b - - 0 79
"Good decisions come from experience, and experience comes from bad decisions."
__________________________________________________________________
Ted Summers
No progress? We have just tested a new bigger net (128x10) and it's 200+ elo stronger. Also v0.5 has been released which brings a huge boost in strenght to the current (and bugged) networks.
Right now the training pipeline is stopped while bugs are fixed and once everything is stable enough, it will continue, probably with the new bigger net.
I'll have to change my gauntlet engines because they are already too weak and "BigNet" (which we call the new enlarged net) crushed like they were toys.
Gauntlet score (last one is the experimental bigger network)
CMCanavessi wrote:
No progress? We have just tested a new bigger net (128x10) and it's 200+ elo stronger. Also v0.5 has been released which brings a huge boost in strenght to the current (and bugged) networks.
Right now the training pipeline is stopped while bugs are fixed and once everything is stable enough, it will continue, probably with the new bigger net.
I'll have to change my gauntlet engines because they are already too weak and "BigNet" (which we call the new enlarged net) crushed like they were toys.
128x10 will probably be retrained, with an improved training schedule. We can expect another strength boost from that, compared to "BigNet" (whoever came up with that name...)
In a few days, we should have a Leela that can compete in the 2300-2500 range...
Laskos wrote:On tactical test suite, also no significant progress. Let's hope for now that move selection in the openings will change this real stalling (that in self-games was artificial). It surely was stuck with some local optima in the openings, and the whole training process was flawed. If not this is the issue, a larger block/filters network will be required.
The change with opening move selection is only going to affect matches, which should yield more accurate self-play Elo ratings in the future. It did not and will not affect the training process at all, since training games always had the full variety in move selection. The most likely candidate for the stall is really the network size, which will be increased within a few days if all goes as planned.
I think that the most candidate for the stall is the fact that LCzero insist to do the same stupid things that alpha zero did instead of trying to improve.
LCzero is weak in seeing mates and they do not consider it as a problem for some reason.
It can miss mate in 1 at least in easy mode.
What about adding a check if there is mate in 1 to LCzero in every position before making playout against itself?
What about simply adding alphabeta to LCzero?
I believe that simple alphabeta when the evaluation at the leaves is simply
expected score at easy mode can do LCzero stronger.
Here is an interesting move-by-move assessment from NM "Jerry" from Chessnetwork, who has played several games with LCZero and plays two games with a recent iteration of the system.
I m sorry, maybe I don't understand it well, but when I look this graph : http://lczero.org/ I don't see progress in elo strenght since many games and days. So perhaps it is not the right thing to watch ?
