TCEC season 20

[Warren D Smith]

Premier Division is going to end later today; at present LcZero and Stockfish are tied for the lead both with 68.18%;

each has 24 wins, 4 losses, and 27 draws so far.  In clear third is KomodoDragon with 58%.

Interesting that these two monsters currently are so equal.  Stockfish has been making major improvements incorporating

and adjusting to having NNUE. What has LcZero been doing? As far as I know nothing except just continuing training?

[Warren D Smith]

Neural net engines are the top 6 in divP; the bottom 2 do not use (or
use little) NNs.
The fastest engine rofchade (100-400 Mnps) is in last place.
Stockfish, currently tied with LcZero for first place, also is pretty fast.
Rofchade appears to be 1.0-1.5X stockfish's speed.
The slowest is stoofvlees with 20-40 Knps.
LcZero (co-leading) also is pretty slow, 20-200 Knps and typically
2-3X stoof's speed.

--
Warren D. Smith
http://RangeVoting.org <-- add your endorsement (by clicking
"endorse" as 1st step)

[hor...@gmail.com]

Afaict - from discord - the thing that the devs have done is encorporate a performance patch for large scale systems.. Apparently it makes lc0 approximately 40% faster on.TCEC hardware.

[glbchess64]

The changes since S19:

- better binary (not for divP but for SuFi), the implementation of some ideas from Ceres results in a x2 speed up. The version number 0.26.3 is an abuse since the binary that plays the SuFi include some patch from future 0.27.0. This version is still under development and will play in S21.

- better parameter tuning for long TC.

- many ideas in training that produce better nets (both in T60 main run and in @jhorthos third party nets).

[Shuo Xiang]

The version playing the SuFi right now is 0.27.0 and not 0.26.3.

So what does "tilps-dje-magic" mean? I especially find the use of the word "magic" hilarious.

Alex

--
You received this message because you are subscribed to the Google Groups "LCZero" group.
To unsubscribe from this group and stop receiving emails from it, send an email to lczero+un...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/lczero/c8d126c7-6a18-42f5-a304-9c21900377b7n%40googlegroups.com.

[glbchess64]

OK, the dev finally change the version number for 0.27.0d. The d means development. In fact, this is not really 0.27.0 but basically 0.23.3 + just a patch for speed. dje (a member of Leela dev team) is the Ceres developer and Tilps (an other member of Leela dev team) adapted an idea from Ceres to Leela that give a x2 to x2.5 speed up. This is idea that is considered "magic". :D

[Shuo Xiang]

Magnifique!

To view this discussion on the web visit https://groups.google.com/d/msgid/lczero/f3c2cddc-3767-4613-9774-bc240ce2a8f5n%40googlegroups.com.

[Warren D Smith]

Lc0-SF superfinal dead even so far. (DivP: also dead even.)

[Warren D Smith]

Stockfish is now 1 game ahead of Lc0 in TCEC 20 superfinal.
Every game pair yielded a tie result except for this one!

Kind of makes you wonder whether we are starting to see the "end of chess" where
both sides play perfectly enough that any match is going to be a tie.
Because at the current rate there are only going to be 2 games in the 100
mitigating against that.

Either that, or the TCEC organizers are going to have to improve their
game-start technology to produce more gamestarts sitting right on the
knife-edge between draw and win... these are probably very difficult
to find, and will keep getting harder to find.
Maybe it will become infeasibly hard to find them.

--
Warren D. Smith

[glbchess64]

The openings of S19 SuFi were too biased and the openings of S20 SuFi not biased enough. In fact divP openings were more biased than SuFi ones! It seems that when there is less than 100cp advantage for SF less than 50cp advantage for Leela these two engines are strong enough to draw. But when the advantage is greater the probability of a 1-0 0-1 result is high.
The opening selecting task for SuFi is very difficult because using Leela or SF to evaluate the opening is not fair (the risk to give an advantage to one these engines is too high).

Watching at the games, we can see  that evaluation of positions are very close so that there is often very long PV agreement and I think this is the main reason for so many draws. (Notice that if SF eval < 150cp Leela eval is about half of SF eval. Not the same scale but the same variations.)

[mjlitola]

I think that TCEC needs much more computational power in the future, so they can play multiple 120 min games in multiple virtual machine which each would have 2080ti or 3090 and 8 or 16-core cpu. SF's development team will sometimes run almost 50.000 games (with draw rate of 82%) to figure out whether code change is improvement or not. It seems that right now the hardware is not strong enough to generate enough games with TCEC current server machine. Low xeon 8 core 2609v4 (passmark of 5970) with 3090 costs 29€ per day, so upgrading to 32 core threadripper would probably be around 40€ per day and you probably need 10 3090 and 10 pcs of TRs to run those 50.000 games, but atleast you would get the winner > it would cost around 3000 euros though, so somebody would need to cover the charge or somehow > I doubt 800 viewers pay that much even if it was on youtube!??

[Warren D Smith]

Maybe "the end of chess" was too pessimistic.
TCEC 20 superfinal now has played 60 games (30 pairs)
with current score SF 4 game-wins, Lc0 3 wins; and of the 30
pairs, 25 were tied, 5 not. The first 16 pairs all were tied, which created
an initial illusion that "chess was ended"; but this was due either to
extremely bad luck or to the organizers intentionally placing all the
more-boring
gamestarts at the beginning. All non-tied pairs were draw+win, there have been
no 2-0 pairs.

[pawel....@gmail.com]

Beginning of TCEC was extremely boring with almost all draws on the board, so I stopped looking there.  To my surprise, when I checked there yesterday the win score was 4:3 in favor of Sf . I ques then, the second part of the match will be more interesting. 

[jmba...@gmail.com]

Bonjour Monsieur

Connaissez-vous un moyen de se procurer ce correctif pour Lc0 ?

Si oui Merci

Jean-Marc

[glbchess64]

" Connaissez-vous un moyen de se procurer ce correctif pour Lc0 ? "

What do you mean by 'corrective' or may be 'fix'? Culd you precise your request?

Qu'ententez-vous par correctif ? Pouvez-vous préciser votre requête.

[Shuo Xiang]

Let me take a stab at translating this:

"Do you know how I could get this patch for Lc0?"

"If yes thanks."

--

You received this message because you are subscribed to the Google Groups "LCZero" group.
To unsubscribe from this group and stop receiving emails from it, send an email to lczero+un...@googlegroups.com.

To view this discussion on the web visit https://groups.google.com/d/msgid/lczero/8ee8ef07-01f1-40ac-9b43-b740dc337264n%40googlegroups.com.

[Warren D Smith]

It turns out (I did not know this originally) that TCEC in its
superfinal now CHRONOLOGICALLY ORDERS all its gamepair starts from
least- to most biased.
(Bias estimated using earlier SF and Lc0 versions, and/or other chess engines.)

This explains all the early tied-gamepairs, and now SF has pulled
ahead of Lc0 by 11 wins to 5 after 90 games.

So it appears that when the gamestart bias is small, Lc0 and SF are
very near equal in strength. When the bias is very large, then of
course they also are equal (game pair result 1-0 then 1-0 the other
way). But for intermediate biases SF is superior to Lc0.

That is the most obvious lesson of the TCEC 20 superfinal. It seems
to point to a flaw in Lc0 net training. You should be playing
training games using biased gamestarts.

[pawel....@gmail.com]

Leela is clearly losing, perhaps her time management is inappropriate. Why not program her only for TCEC to make the first move after some set time, such as 15 minutes after finishing the book?  The reason is that the first few moves after finishing the book are crucial for the game to develop properly, and my intuition tells me that she should take more time to properly evaluate the position, right after book end. If you look at a few games, you'll see that she doesn't think too long on the first move, maybe this should be manually addressed by forcing her to think longer? This should be easily coded, e.g. after finishing the book, don't think less than 10 minutes on the first move. Later in the game, she spends more time, but the position is already lost, so why is she thinking so long? There is no reason, the position has already been lost!  

[glbchess64]

There is some controversy on time management. But it seems not so bad. Leela has difficulties with 10" increment at end of game. To avoid tactical blunder, she need some time. Especially if SF plays an unexpected move. In such a case, she needs to rebuild the search tree. Often the two engines agree (no tree to rebuild) but in complex positions where one of them can win the move of the opponent is sometimes a surprise and the tree have to be rebuilt.