My test for Kayra 1.0 v6 bmi2

[AlexChess]

I'm just updating Raubfisch X48b1 on my SuperBlitz... too.

Until now X47e1 was not on top 10 (Kayra 1.0 v6 is much stronger)

New results soon, but not today

Best regards, Alex

[AlexChess]

There is a serious gap between Kayra 1.0 v6 bmi2 and the Raubfisch, Fisherov, M31 agent & co team. I will wait for a new development version of Stockfish to appear this weekend ......

You don't need to wait 'til the weekend, as it is nearly the same as SF dev with the new net from Dec. 13.
https://abrok.eu/stockfish/
The 'author' just randomly changed some psqt (which most of time are not even used...because of nnue).
I checked your first game and it plays identically.

It is just another illegal SF clone, which now even demands to be paid via patreon at that ugly/unbearable engines diary site.

The parameters changed in the Evaluation section:256
The parameters changed in the PSQT section:416
The parameters changed in the Pawns section:12
Total parameter changes:684.

All of the changes are radical changes. I didn' t make any minor changes in the codes. Psqt codes are very important codes. In the test I made yesterday by giving strange values to the psqt codes, this engine was defeated by Stockfish with a 100-0 score.
I have not asked anyone for money for this engine and I never will. I just sent this engine some testers to test.This engine is still at the beginning of its development process and no doubt that in 2022 Kayra will be an open source code chess engine.

+1

[AlexChess]

Download:
https://filehorst.de/d/egFrDjsy

THANK YOU ! SUPER !

Using the .pgn Kayra matches collection you can made (easy) an opening book in .abk format under Arena GUI. Here is a short tutorial:

https://www.chess2u.com/t4215-how-to-ma ... p-tutorial

This book can be used like GUI book for all engines in matches/tournaments in Arena GUI !

I'm wondering why, if some original authors here know so well Stockfish code, their engine Is still 300 ELO worse, while with (supposed) random changes Keyra 1.0 v6 is so strong according to most testers ...Mehmet and Sylwy you are welcome on BanksiaGui forum to test all SF derivatives that you likes, skipping boring critics.
https://banksiagui.com/forums/viewtopic.php?t=19

Best regards, Alex

[connor_mcmonigle]

he compiled just his psqt changes (which have near zero effect due to nnue, but this was already told so often to no avail)

I wrote before
"Psqt codes are very important codes. In the test I made yesterday by giving strange values to the psqt codes, this engine (Kayra X NNUE ) was defeated by Stockfish NNUE with a 100-0 score."


Differences between Stockfish and Kayra codes.
https://github.com/Vizvezdenec/Stockfis ... baadf0a077

But this list is incomplete. They forgot to add some changes.

You seem to be confused. It's trivial to make SF massively weaker by choosing awful PSQT values as Stockfish switches to the classical evaluation function when handling winning/lost positions as decided by the classical evaluation function. However, disabling the classical evaluation function entirely only loses 1-2 elo. The classical evaluation function is irrelevant to Stockfish's overall strength (only 1-2 elo can be attributed to it).

1)I read in the stockfish forum a claim that it is 10 elo and not 1-2 elo

https://groups.google.com/g/fishcooking/c/v5cqRoHaetk

I do not know if it is correct.

2)Even if it is 1-2 elo it only means that the classical evaluation of today does not change the playing strength of stockfish by more than 1-2 elo.
It does not mean that the classical evaluation function is irrelevant because of the following reasons:

a)maybe by some change in the classical evaluation you can get 10 elo improvement in stockfish.

b)elo is not everything.
If stockfish can find mates faster in positions when one side has a big material advantage then it is important for people who use it for analysis and if I understand correctly with a big material advantage stockfish does not use NNUE but only classical evaluation.

I found in the past that stockfish is slow in mating the opponent with black when white starts without queen d1.
If you can make it faster relative to other engines thanks to improving the classical evaluation then I consider it as more important improvement for users then adding 5 elo.

If that 10 Elo figure is to be believed, then perhaps, due to some tweaks to NNUE scaling or SPSA tuning, Stockfish grew dependent on the classical evaluation function in some nontrivial way.

In any case, if you read through the changes, you'll see that a bunch of the PSQT values are left uninitialized... It proved a clear -10 Elo regression at FishTest, likely due to the high number of mostly zero valued uninitialized classical eval terms result in the classical eval threshold almost never triggering.

The classical evaluation function is only used in decided positions -> there is very minimal Elo to be gained here, almost certainly far less Elo than the 10 Elo you suggest is possible.

[connor_mcmonigle]

Download:
https://filehorst.de/d/egFrDjsy

THANK YOU ! SUPER !

Using the .pgn Kayra matches collection you can made (easy) an opening book in .abk format under Arena GUI. Here is a short tutorial:

https://www.chess2u.com/t4215-how-to-ma ... p-tutorial

This book can be used like GUI book for all engines in matches/tournaments in Arena GUI !

I'm wondering why, if some original authors here know so well Stockfish code, their engine Is still 300 ELO worse, while with (supposed) random changes Keyra 1.0 v6 Is so strong according to most resters ...Mehmet and Sylwy you are welcome on BanksiaGui forum to test all SF derivatives that you likes skipping boring critics.
https://banksiagui.com/forums/viewtopic.php?t=19

Best regards, Alex

You really are not the brightest.

Weakening Stockfish with some silly random changes is easy.
Writing an original engine, even one -300 Elo to Stockfish, is very difficult.

This shouldn't be difficult to comprehend.

[mehmet123]

I'm just updating Raubfisch X48b1 on my SuperBlitz... too.

Until now X47e1 was not on top 10 (Kayra 1.0 v6 is much stronger)

New results soon, but not today

Best regards, Alex

I was happy that Kayra used the NNUE net effectively as I expected. Although Kayra is at a very low level like Stockfish 1.0 without NNUE net, it has achieved very good results using NNUE especially rapid and long time control games.
At Kayra 1.1 dev I especially focused on the search part. There is no point in using parameters except that are compatible with NNUE net. Thanks to Kayra parameters I got important information about NNUE that I want to learn.
https://sites.google.com/site/computers ... ct-nn-cs-8

[AlexChess]

People still using raubfisch? Disgusting

Dear Sopel,
excepted for the GPLv3 infringement (I agree with you Is really bad) why are you disgusted by Raubfisch? It's only a 1:1 SF clone? I respect your option and I can consiider to remove It from my SuperBlitz.

Best regards, Alex

[AlexChess]

Due to internet connection coming and going, the other message was sent twice by mistake.

Thanks Sylwy for the test. This is the first time I see such a big difference between two Stockfish derivative chess engines but I don't think Kayra is that strong yet. Because development process is new and still continues. I believe that some techniques I have found today will carry Kayra to better places in the future.
My first goal was to make radical changes into the Stockfish codes and I think I've done that more than enough in such a short time. Radical changes in 684 parameters is proof of this. I didn' t make any minor changes to any of these 684 parameters.
As I always say, I can't say that Kayra is the most powerful chess engine, but I am sure that Kayra is the engine that has the less similarity with Stockfish among the Stockfish derivative chess engines.

If Stockfish derivatives would be only 1:1 clones, the crosstable results would be an SF 1:1 copy, too. I think that also fine tuning engines to gain some extra points is very hard and requires programming and chess skills. Obviously not like writing a new engine from scratch, but isn't trivial. And most "original" authors start from an existing engine, like Stockfish Team with Glaurung. We cannot reinvent the wheel each time, and modify GPLv3 or MIT sources is absolutely allowed. I deprecate only the few authors that do not release their modified code to help programmers community to grow.


(sorry for wrong rankings, table is ordered by scores)

[Uri Blass]

he compiled just his psqt changes (which have near zero effect due to nnue, but this was already told so often to no avail)

I wrote before
"Psqt codes are very important codes. In the test I made yesterday by giving strange values to the psqt codes, this engine (Kayra X NNUE ) was defeated by Stockfish NNUE with a 100-0 score."


Differences between Stockfish and Kayra codes.
https://github.com/Vizvezdenec/Stockfis ... baadf0a077

But this list is incomplete. They forgot to add some changes.

You seem to be confused. It's trivial to make SF massively weaker by choosing awful PSQT values as Stockfish switches to the classical evaluation function when handling winning/lost positions as decided by the classical evaluation function. However, disabling the classical evaluation function entirely only loses 1-2 elo. The classical evaluation function is irrelevant to Stockfish's overall strength (only 1-2 elo can be attributed to it).

1)I read in the stockfish forum a claim that it is 10 elo and not 1-2 elo

https://groups.google.com/g/fishcooking/c/v5cqRoHaetk

I do not know if it is correct.

2)Even if it is 1-2 elo it only means that the classical evaluation of today does not change the playing strength of stockfish by more than 1-2 elo.
It does not mean that the classical evaluation function is irrelevant because of the following reasons:

a)maybe by some change in the classical evaluation you can get 10 elo improvement in stockfish.

b)elo is not everything.
If stockfish can find mates faster in positions when one side has a big material advantage then it is important for people who use it for analysis and if I understand correctly with a big material advantage stockfish does not use NNUE but only classical evaluation.

I found in the past that stockfish is slow in mating the opponent with black when white starts without queen d1.
If you can make it faster relative to other engines thanks to improving the classical evaluation then I consider it as more important improvement for users then adding 5 elo.

If that 10 Elo figure is to be believed, then perhaps, due to some tweaks to NNUE scaling or SPSA tuning, Stockfish grew dependent on the classical evaluation function in some nontrivial way.

In any case, if you read through the changes, you'll see that a bunch of the PSQT values are left uninitialized... It proved a clear -10 Elo regression at FishTest, likely due to the high number of mostly zero valued uninitialized classical eval terms result in the classical eval threshold almost never triggering.

The classical evaluation function is only used in decided positions -> there is very minimal Elo to be gained here, almost certainly far less Elo than the 10 Elo you suggest is possible.

How do you define if a position is a decided position?
If the classical evaluation does it(for example a rule like if the classical evaluation is bigger than 5 pawns then the position is decided position otherwise switch to NNUE) then maybe you can earn more by improving the classical evaluation and decide that more positions are defined to be decided positions(without improving the classical evaluation it may be too risky to decide that more positions are defined to be decided).

[mehmet123]

If Stockfish derivatives would be only 1:1 clones, the crosstable results would be almost the same. I think that also fine tuning engines to gain some extra point is very hard and required programming and chess skills.

I can't agree with you on this issue, the number of games is not too many and there is no big difference in power between these engines. If you add a chess engine in this list with a different name, the performance of this engine will likely be different. Unfortunately, it is very difficult to say that a chess engine is not a clone by looking at this table