New engine releases & news 2021

[kranium]

I don't understand why Norm would suggest to CEGT to combine 8 and 8.1 results when the evaluation per Ed's SIM tests are very very different.
OK I'm not convinced myself on SIM tests such as those but many people more expert than me are.

In my testing, they test identically strength-wise, hence my recommendation...to save time and effort.
The sim depth=1 difference is likely due to the eval changes I made, listed under history on web site.

Concerning sim depth=1
I believe Thomas Miller is at least partially correct about pruning. I'm convinced it's closely related to lazy eval, as I've mentioned numerous times in engine origins as criticism for using depth=1 as a reliable tool to measure engines by.

Particular implementations are likely to register high at Depth=1, Komodo? Schooner?
but would test much more normal when used with movetime=100 for ex., which tests the whole engine.

[kranium]

The complication is that Ed says it is easy for authors to manipulate the SIM-depth=1 test.

I believe defeating the sim Depth=1 test purposefully would be difficult, unless the download from Ed's site allows running depth=1, otherwise there's no way for to test. I believe the version I have from quite some time ago, only tests using 'movetime'.

[Rebel]

The complication is that Ed says it is easy for authors to manipulate the SIM-depth=1 test.

I believe defeating the sim Depth=1 test purposefully would be difficult, unless the download from Ed's site allows running depth=1, otherwise there's no way for to test. I believe the version I have from quite some time ago, only tests using 'movetime'.

Ed doesn't offer any Fire for download.

Your Fire 8.0 runs the sim-test at depth=1 just fine.

go depth 1 results:
Fire 7.1 - bestmove e2e4

Fire 8.0

Code: Select all

Fire 8 x64 popcnt
go depth 1
info time 1 nodes 1 nps 1000 tbhits 0 depth 1 seldepth 2 multipv 1 score cp -35 pv a2a3
info time 1 nodes 2 nps 2000 tbhits 0 depth 1 seldepth 2 multipv 1 score cp -1 pv b2b3
info time 2 nodes 3 nps 1000 tbhits 0 depth 1 seldepth 2 multipv 1 score cp 105 pv c2c3
info time 2 nodes 20 nps 10000 tbhits 0 depth 1 seldepth 2 multipv 1 score cp 105 pv c2c3
bestmove c2c3

1.c3 best move with 1.05 score, seriously?

Fire 8.1

Code: Select all

Fire 8.1 x64 popcnt
info string 40 available cores
go depth 1
info time 1 nodes 1 nps 1000 tbhits 0 depth 1 seldepth 2 multipv 1 score cp -2 pv a2a3
info time 1 nodes 2 nps 2000 tbhits 0 depth 1 seldepth 2 multipv 1 score cp 0 pv b2b3
info time 1 nodes 3 nps 3000 tbhits 0 depth 1 seldepth 2 multipv 1 score cp 10 pv c2c3
info time 2 nodes 4 nps 2000 tbhits 0 depth 1 seldepth 2 multipv 1 score cp 27 pv d2d3
info time 2 nodes 5 nps 2000 tbhits 0 depth 1 seldepth 2 multipv 1 score cp 42 pv e2e3
info time 2 nodes 20 nps 10000 tbhits 0 depth 1 seldepth 2 multipv 1 score cp 42 pv e2e3
bestmove e2e3

That looks normal.

[kranium]

I believe defeating the sim Depth=1 test purposefully would be difficult, unless the download from Ed's site allows running depth=1, otherwise there's no way for to test. I believe the version I have from quite some time ago, only tests using 'movetime'.

Ed doesn't offer any Fire for download.

Your Fire 8.0 runs the sim-test at depth=1 just fine.

I not suggesting you offer Fire for download!
...and I'm well aware Fire runs at depth=1, you've misunderstood.

I don't believe your download of sim has the capacity to run Depth=1

[Guenther]

Fire 8.1 with source code has been released. https://github.com/FireFather/fire/releases/tag/8.1

You might also be interested in this: http://talkchess.com/forum3/viewtopic.p ... 9&start=80

I'm glad Norman Schmidt has finally made Fire a GPL-compliant engine and made his source code available to the public.

I reported to the moderation team about hi-jacking my thread.

Starting with this post my thread was hi-jacked for Fire 8 origins (which is ofc a SF clone).
Only one post about another new program done by me should remain there!

So please split my thread about new engine releases in two parts!

Thanks

[Rebel]

I believe defeating the sim Depth=1 test purposefully would be difficult, unless the download from Ed's site allows running depth=1, otherwise there's no way for to test. I believe the version I have from quite some time ago, only tests using 'movetime'.

Ed doesn't offer any Fire for download.

Your Fire 8.0 runs the sim-test at depth=1 just fine.

I not suggesting you offer Fire for download!
...and I'm well aware Fire runs at depth=1, you've misunderstood.

I don't believe your download of sim has the capacity to run Depth=1

SIMEX output first 5 (of 8238) positions of Fire 8.0

Code: Select all

3q1rk1/3n1ppp/1p1p1b2/1R1P4/8/r1pNP1P1/P1Q2PBP/5RK1 w - - bm h3; ce -4; acd 1;
r3k2r/pp1qppbp/2np1np1/8/2PNP3/2N1BP2/PP4PP/R2QK2R b KQkq - bm Ng4; ce -8; acd 1;
r6r/1b3pk1/1qp1p2b/p1NnP2p/Pp4pP/1Q4B1/4BPP1/2RR2K1 w - - bm Qf3; ce -164; acd 1;
r2qr1k1/1bp1bppp/p1np1n2/1p2p3/P3P3/1B1P1N2/1PPN1PPP/R1BQR1K1 w - - bm h3; ce -6; acd 1;
rnbq1rk1/ppp1bpp1/4pn1p/3p4/2PP3B/2N1PN2/PP3PPP/R2QKB1R b KQ - bm Nc6; ce 268; acd 1;

As you can see depth=1.

The first position is already big fun. Running Fire 8 in a GUI it correctly plays h3 with a -4 score at depth=1



Then repeating takeback + force move it starts to play all kind of different moves and scores (see the orange moves) even to the point of giving away a queen with a 3 pawns plus score. Of course you took the obfuscation out of 8.1 and 8.1 keeps on playing f4 with a -7 score as in the output of SIMEX.

Code: Select all

3q1rk1/3n1ppp/1p1p1b2/1R1P4/8/r1pNP1P1/P1Q2PBP/5RK1 w - - bm f4; ce -7; acd 1;

On the internet truth always come out.

[kranium]

Then repeating takeback + force move it starts to play all kind of different moves and scores (see the orange moves) even to the point of giving away a queen with a 3 pawns plus score. Of course you took the obfuscation out of 8.1 and 8.1 keeps on playing f4 with a -7 score as in the output of SIMEX.

Code: Select all

3q1rk1/3n1ppp/1p1p1b2/1R1P4/8/r1pNP1P1/P1Q2PBP/5RK1 w - - bm f4; ce -7; acd 1;

On the internet truth always come out.

So basically you played it enough, forced it to moved back and forth 7- 8 times until finally you found something that suits your agenda to post.
Well I do know one thing...if anyone does indeed defeat your sim depth 1 testing, I would certainly understand.

You manipulate methods until you get results to look as controversial as possible, and use it and bold red-letter innuendo on you site to not-so-subtlety discredit others including Komodo, Shredder, Schooner, and many more.

My apologies to Guenther for answering posts in this thread.

[Rebel]

Maybe you can fool a couple of people, but not the engine programmers.

What a waste of your talent!

[pedrox]

DanaSah 8.8 allows the use of a NNUE type neural network. It includes a network that has been created from another network created with CCRL 40/2 positions. More information in the readme.txt file. Although there are many changes, the version with the classic evaluation does not gain strength, about the same as version 8.3. The NNUE version with the dananet1 network win until 500 points, depending on the executable used (would gain up to about 600 with a stockfish net).

https://www.mediafire.com/file/a4bq4vxt ... 8.zip/file

The network ccrl404_fn.nnue is created with all positions from the CCRL 40/2 list database. The games were converted to bin format and do not include evaluation and depth of the engines as the pgn database does not contain comments. For this reason the network is trained to learn by results. The network reminds me of the early lc0 networks where the game is very positional and less tactical. The network is not very strong, about 425 Elo points less than stockfish master, but it has possibilities to use it for new data and make it stronger.

https://www.mediafire.com/file/e7tjxfzh ... n.zip/file

There is a version 8.8 with web gui with limited strength (evaluation classic):

http://danasah.pythonanywhere.com/

So which of the 2 is "more original" or "more pure danasah"? Which one should we use for tournaments?

The default network for this version of danasah is dananet1.nnue. This is the version to test, it does not use evaluations from stockfish or other engines. He has learned by first training by results with ccrl 40/2 positions and then with depth 5 positions evaluated with the same net to win tactically.

The following image shows the ccrl402 network. It can be seen that the network does not yet have a great knowledge of the value of the pieces and bases all its play on positional evaluation. Click for a better view of the image.



In dananet1 it can now be seen that the pieces already have much more weight in the evaluation, he will be much stronger tactically and gained about 300 Elo points.



I think the dananet1 network could be used for my engine and derivatives. And the ccrl402 network in a more free form, could be used for testing and as a basis for others. What I have to say is that these networks have little to do with classical evaluation, which I understand as normal.

If Ed would like to test the networks with SIM, I would be happy to see the results.

[Modern Times]

DanaSah 8.8 allows the use of a NNUE type neural network. It includes a network that has been created from another network created with CCRL 40/2 positions. More information in the readme.txt file.

https://www.mediafire.com/file/a4bq4vxt ... 8.zip/file

Does DanaSah 8.8 play chess960 under Cutechess GUI, either as UCI or WinBoard ?