News about Rebel or Chess System Tal ?

I wish I had new that, ProDeo plays much too fast on sudden depth levels and I never cared, tested.

Does CCRL also plays sudden depth?

This I can not say, but I think that this is the case, at least some testers, I think that Graham can give an answer.

This is the case with Ipman and Stefan Pohl, i.e. they test with time controls such as 3 minutes +1 second....

What gives more elo ? Pondering or better timing in blitz levels ? I would say ponder gives more elo in ANY time control, no matter if blitz , average or tournament.

Only test suite and testing with ponder OFF gives no better results.

it's hard to say, most rating tests are done without Ponder. Ponder plays a role in official tournaments. A better time control is in any case a higher gain than maybe 5 percent speed advantage due to a better compilation...

In a test with Ponder I think Lc0 should have a certain advantage, because it has the complete search tree in memory and just continues to calculate, but I am only a "prevented" Lc0 expert, i.e. I don't understand the MCTS search and the differences to the search of an AB engine until today... (why no quiescent search ?)

If the tests are on one PC (most are, I assume), then ponder is going to wreak havoc with the allocation of memory and cpu resources, such that one engine can unfairly gain against the other. I always assumed ponder=off to make even playing field.

When i have 8 cores each engine gets 4 cores. With the 16 core machine each engine 8 cores.
I find pondering natural. Its what humans do too.
Not pondering looks like wasting time.
Only the very early dedicated chess computers of 1978-79 did not ponder.
MkV (Broughton) pondered, mephisto 1 did (Nitsche+Henne) fidelity sensory 9 (Spracklen) did.
Superchess3.5 did not ponder but clockchess89 did.

for private tests Ponder makes sense, for tests for rating lists it makes more sense to use the existing threads for significantly more games in the same time....

Peter Berger wrote:

Admin wrote:

Oh well, forgot to tell, PB still under construction.

Yes, use the Noomen book, no update needed.

Or don't .

Welcome Rebel16 on my computers.

My knowledge in this line is heavily outdated (like 19XX) , so for me 12. h3 is kind of a novelty. I didn't look things up, so who knows .

It is one game, but though Crafty loses some, even Stockfish rarely delivers such strong blows against it - impressive.

But what I want to give an applause for, is the release style-wise. We get an executable, we get one parameter, we can adjust threads. THAT's the spirit - just as in the old days where programmers were expected to work out optimal settings for themselves    



[Event "Lang 120min+10sek"]
[Site "Berlin"]
[Date "2022.11.30"]
[Round "?"]
[White "Rebel 16"]
[Black "Crafty 25.6"]
[Result "1-0"]
[ECO "C82"]
[PlyCount "75"]
[TimeControl "7200+10"]

{4096MB, LAPTOP-NCDN8BTK} 1. e4 {[%eval 50,28] [%emt 0:04:58]} e5 {[%emt 0:00:
07]} 2. Nf3 {[%eval 61,28] [%emt 0:01:33]} Nc6 {[%emt 0:00:08]} 3. Bb5 {
[%eval 58,27] [%emt 0:01:25]} a6 {[%emt 0:00:07] (Sf6)} 4. Ba4 {[%eval 71,28]
[%emt 0:01:31]} Nf6 {[%emt 0:00:08]} 5. O-O {[%eval 66,28] [%emt 0:01:22]} Nxe4
{[%emt 0:00:08] (b5)} 6. d4 {[%eval 64,29] [%emt 0:01:30]} b5 {[%emt 0:00:06]
(exd4)} 7. Bb3 {[%eval 95,30] [%emt 0:01:20]} d5 {[%emt 0:00:07] (Le7)} 8. dxe5
{[%eval 82,30] [%emt 0:01:43]} Be6 {[%emt 0:00:09]} 9. Nbd2 {[%eval 51,30]
[%emt 0:02:44]} Nc5 {[%emt 0:00:08]} 10. c3 {[%eval 58,30] [%emt 0:01:08]} Bg4
{[%emt 0:00:08] (d4)} 11. Bc2 {[%eval 127,30] [%emt 0:03:49]} Qd7 {[%emt 0:00:
08] (Le7)} 12. h3 {[%eval 183,30] [%emt 0:03:19]} Bh5 {[%emt 0:03:08] (Lxf3)}
13. b4 {[%eval 137,30] [%emt 0:01:41]} Ne6 {[%emt 0:01:14] (Sa4)} 14. a4 {
[%eval 163,29] [%emt 0:01:17]} Na7 {[%emt 0:11:16] (Tb8)} 15. Nb3 {[%eval 213,
30] [%emt 0:03:28]} Be7 {[%emt 0:07:12] (bxa4)} 16. Nbd4 {[%eval 199,30] [%emt
0:02:43]} Nxd4 {[%emt 0:02:30]} 17. Qxd4 {[%eval 201,29] [%emt 0:00:01]} Bxf3 {
[%emt 0:02:15]} 18. gxf3 {[%eval 198,28] [%emt 0:00:01]} O-O {[%emt 0:02:22]}
19. Kg2 {[%eval 227,28] [%emt 0:00:01]} Qe6 {[%emt 0:03:13] (Tfd8)} 20. f4 {
[%eval 310,28] [%emt 0:01:12]} f5 {[%emt 0:04:36]} 21. Qd3 {[%eval 328,28]
[%emt 0:00:01]} c6 {[%emt 0:02:41] (Kh8)} 22. Be3 {[%eval 304,32] [%emt 0:03:
30]} Kh8 {[%emt 0:00:09]} 23. Kh2 {[%eval 304,31] [%emt 0:00:53]} g6 {[%emt 0:
02:25] (bxa4)} 24. Rfc1 {[%eval 359,29] [%emt 0:03:27]} Qc8 {[%emt 0:04:33]
(Kg7)} 25. Bd1 {[%eval 409,29] [%emt 0:03:19]} Qb7 {[%emt 0:01:47] (De6)} 26.
Bf3 {[%eval 364,30] [%emt 0:02:53]} Kg7 {[%emt 0:10:16] (Sc8)} 27. Bd4 {
[%eval 426,25] [%emt 0:02:30]} Rfc8 {[%emt 0:04:23] (bxa4)} 28. a5 {[%eval 460,
27] [%emt 0:02:59]} Bh4 {[%emt 0:00:50] (Tg8)} 29. e6+ {[%eval 469,27] [%emt 0:
00:57]} Kh6 {[%emt 0:03:15]} 30. Re1 {[%eval 526,26] [%emt 0:00:01]} Rf8 {
[%emt 0:05:35] (Tg8)} 31. e7 {[%eval 814,27] [%emt 0:00:56]} Bxe7 {[%emt 0:01:
27]} 32. Re6 {[%eval 747,26] [%emt 0:00:01]} Qc7 {[%emt 0:01:06] (Dd7)} 33. h4
{[%eval 1050,25] [%emt 0:02:04]} Rf7 {[%emt 0:02:30] (Lxh4)} 34. Kg2 {[%eval
1510,25] [%emt 0:02:00]} Raf8 {[%emt 0:02:59]} 35. h5 {[%eval 1811,27] [%emt 0:
00:01]} Bf6 {[%emt 0:01:48] (Tg8)} 36. hxg6 {[%eval 1971,27] [%emt 0:00:53]}
hxg6 {[%emt 0:00:47]} 37. Qe3 {[%eval 2042,28] [%emt 0:00:58]} Kg7 {[%emt 0:00:
57] (Dd8)} 38. Bxf6+ {[%eval 2321,28] [%emt 0:00:52]} 1-0

yeah h3, is a rare move kinda, because Bg4 was a massive sideline, like +1 under stockfish

Dio wrote:

for private tests Ponder makes sense, for tests for rating lists it makes more sense to use the existing threads for significantly more games in the same time....

Why not "doing one without neglecting the other"?
http://www.cegt.net/rating5plus3pbon.htm
~420.000 games

Chris Whittington wrote:

If the tests are on one PC (most are, I assume), then ponder is going to wreak havoc with the allocation of memory and cpu resources, such that one engine can unfairly gain against the other. I always assumed ponder=off to make even playing field.

With modern computers, modern GUI etc., I do not think this is an issue any more. Maybe on one or two core machines 15 or 20 years ago. I am - carefully worded - a bit shocked about some statements here concerning Ponder.  

http://www.cegt.net/rating5plus3pbon.htm
~420.000 games
ALL played on one PC. No one except SSDF plays with Ponder on two machines nowadays with Auto232 cable or other oldfashioned "solutions" from the Stone Age.

Currently most tests for this list are played on my Ryzen 9 3950x (16 Cores, 32 Threads) with 32GB RAM and BanksiaGUI.
Concurrency is set to 6 or 7, i.e. 12 to 14 Cores are busy playing one tournament, gauntlet or match. Maybe one also could play with higher concurrency, but I don't like that. Results are quite consistant compared to our PonderOFF lists.
Btw: Rebel 15.1 runs perfectly with Ponder. 1500 games, not a single problem/crash/time forfeit

Another example, much more extreme, was Ingo Bauers "IPON".
Up to his retirement from computerchess he played hundredthousands of games with 5 or more Hexacore PCs simultaneousely over Network/LAN.
Never heard about fundamental technical (!) problems. There were some discussions about his conditions, e.g. that he never published his opening suite. But this does not belong here.

It's of course up to you to implement ponder or not.
I would appreciate it very much

Best
Wolfgang
CEGT Testing Group

Mclane wrote:

When i have 8 cores each engine gets 4 cores. With the 16 core machine each engine 8 cores.
I find pondering natural. Its what humans do too.
Not pondering looks like wasting time.
...

Agreed...

Judging from the Rebel 16 release date, CSTal is going to be a whole Christmas present

Chris Whittington wrote:

If the tests are on one PC (most are, I assume), then ponder is going to wreak havoc with the allocation of memory and cpu resources, such that one engine can unfairly gain against the other. I always assumed ponder=off to make even playing field.

At least on the computers I have built. This has never been a issue. And this is easily checked by the stats for each chess engine produced by the GUI. To see that ponder on is 100% ok if done correctly on a single computer.

Meaning you have to split the CPU cores between the 2 chess engines and the memory. And if you have at least 16 CPUs. This is my preferred way of testing. As most chess engines gain very little going from 8 to 16 cores.

This is not true for some chess engines like Dragon 3.1 using MCTS. As that search scales very well with more cores.

I could see if you have a prebuilt with 1 stick of memory as many do to save money. This could be a issue using ponder on. Using a single computer system.

SSDF has always used ponder on to generate their rating list. And that use to be the testing standard.

And if you have 2 computer systems with equal specs. Banksia GUI allows you to connect the computers together using wifi. And you can test like SSDF if you wish. As another testing option with ponder on.

And you could run many games at once with many chess engines. On my system I could play 16 chess engines using 1 core each with ponder on that also works just fine.

If you doubt this. Just try it. It works perfectly on a Threadripper CPU.

pohl4711 wrote:

I had no freezes, but something is definitly wrong with Rebel. Look at move 109 by Rebel 16 (black)...
Drawn position, but a thinking depth of 1 after 1.2 seconds and a bogus evaluation: Rh6+ {-M124/1 1.2s}
(This happens again and again in my testrun) (I use cutechess-cli)

We checked this, likely a display glitch, in any case without consequences for playing strength.

pohl4711 wrote:

(And Rebel 16a moves still too fast. Now, 7min 5secs per game in my testings (better than Rebel 16 (6min 50sec))
Normal average value in my ratinglist is 7min 45secs

Working on this.

Apparently the world has changed since we both retired during the                   -         

Rebel 16 avx2 added to list ,played 23.000games TC: 10s+1s ,more details on website.

https://ipmanchess.yolasite.com/i9-7980xe.php

Rebel 16 did it well ,first time it could enter this list ,even with the known problems as above..

On other system it's playing 1200games @5min.+0s ,results will follow..

Ipman.

Ipman thanks.

I am pretty sure the current best version "Rebel-16a" will do a lot better on the time levels you test.

Eventually there will be a final update with a tuned sudden death level.

Meanwhile on normal time control (CCRL 40/15) results are pretty good sofar as I can judge it.

https://tlcv.net/16070/result-table [1 cpu]
https://tlcv.net/16061/result-table [4 cpu]

Dio wrote:

In a test with Ponder I think Lc0 should have a certain advantage, because it has the complete search tree in memory and just continues to calculate, but I am only a "prevented" Lc0 expert, i.e. I don't understand the MCTS search and the differences to the search of an AB engine until today... (why no quiescent search ?)

That is true for Lc0, and Dragon MCTS but what is also true for Lc0, and Dragon MCTS is that the tree is not lost with ponder off. And Lc0 and Dragon MCTS will use that data again on the next move. How much of that tree is left for Lc0 and Dragon MCTS to use depends on the move played by the other chess engine. But that is exactly the same for ponder on with Lc0 and Dragon MCTS.

And this video could answer your questions on the MCTS.



I just noticed that Larry agrees that using more then 8 cores with NNUE is pretty useless. Nice!

Status for Chess-System-Tal-1.51:

written new code for time control - being tuned and tested at the moment. Written code for ponder - awaits testing. Might have fixed -M124 weirdness - difficult to say, since I can't get it to repeat.

All needs porting back to Rebel-16a

where are the test versions ?

Rebel 16 avx2 also added in other list ,played 1200games @5min+0s ,more details on website

https://ipmanchess.yolasite.com/i7-11800h.php

Gives +81Elo above Rebel 15.1 !!

Ipman.

Ipmanchess wrote:

Rebel 16 avx2 also added in other list ,played 1200games @5min+0s ,more details on website

https://ipmanchess.yolasite.com/i7-11800h.php

Gives +81Elo above Rebel 15.1 !!

Ipman.

+ 81 elo?? There is no need to retire Rebel just yet Ed... Maybe you can add another 80-100 elo so it can join the top 10 on the list?

Damir Desevac wrote:

Ipmanchess wrote:

Rebel 16 avx2 also added in other list ,played 1200games @5min+0s ,more details on website

https://ipmanchess.yolasite.com/i7-11800h.php

Gives +81Elo above Rebel 15.1 !!

Ipman.

+ 81 elo?? There is no need to retire Rebel just yet Ed... Maybe you can add another 80-100 elo so it can join the top 10 on the list?

Most of the elo comes from the switch to the CSTAL search.

Damir Desevac wrote:

Ipmanchess wrote:

Rebel 16 avx2 also added in other list ,played 1200games @5min+0s ,more details on website

https://ipmanchess.yolasite.com/i7-11800h.php

Gives +81Elo above Rebel 15.1 !!

Ipman.

+ 81 elo?? There is no need to retire Rebel just yet Ed... Maybe you can add another 80-100 elo so it can join the top 10 on the list?

stockfish and major engines see minimal elo gains of maybe 1 elo per patch, and they have things like FishTest where they can quick check patches and see if they're good, also they incorporate 960 games into the net which I'm guessing Rebel and CSTal don't because they don't support it ATM. I'm not en engine dev so I don't know how adding a couple hundred thousand 960 games would affect it in openings