Koivisto 7.5 - Fire 8.3

[Koivisto]

https://ibb.co/jfLNt4y
It also scores the startposition quite high compared to most engines.

The Koivisto nnue eval is multiplied by a factor of 2 when all pieces are on the board compared to all pieces off the board. (1.6x for early game - 0.8x late game) (scaled by phase). This is done in a lot of engines, although I think we are the most agressive.

[Jouni]

Koivisto is interesting new engine. Very strong even if exe is below 3 Megabytes. But testsuites are not it's strong point. In many suites like Hardtalkchess it hardly solves any position. Maybe some 20/116.

[Koivisto]

Koivisto is interesting new engine. Very strong even if exe is below 3 Megabytes. But testsuites are not it's strong point. In many suites like Hardtalkchess it hardly solves any position. Maybe some 20/116.

that is sort of expected. Our reduction & research conditions sort of try to make sure we are not wrong rather than finding moves that might actually be really good, conceptually inspired by "worst move observation". If it's garbage or not is questionable, but at least it has all gained at some point in Koi development (aka when it was added ).

For example our iterative researches for root moves should make it harder to find a new strong bestmove (more chances to fail), but if a move manages to get through all the iterations without failing once, its much better tested than it would be otherwise.

[AndrewGrant]

Ethereal vs Koivisto: 2107 - 1686 - 9319 [0.516] 13112

./cutechess -repeat -recover -variant standard -resign movecount=3 score=400 -draw movenumber=40 movecount=8 score=20 -concurrency 32 -games 64000 -engine cmd=Ethereal proto=uci tc=60.00+0.60 option.Hash=64 name=Ethereal -engine cmd=Koivisto proto=uci tc=60.00+0.60 option.Hash=64 name=Koivisto -openings file=4moves_noob.epd format=epd order=random -pgnout games.pgn -tb /home/andrew/Syzygy/

Ethereal 13.35, with current best net. Koivisto "half_network_1024_2" branch, with suggested Network, as of 11/16/2021.

The machine running the games is a Ryzen 1950x. This has AVX2 support, however it does not have 256-bit support, so the AVX2 support is more akin to SSSE3/SSE4 when doing integer operations. I believe this hurts Ethereal more than Koivisto. However, Koivisto has AVX512 support that Ethereal does not, and tests from another user show the results inverted when testing Koivisto AVX512 vs Ethereal AVX2.

[Madeleine Birchfield]

Koivisto most powerful and interesting engine that I see currently out there.
Surely its ahead of E13.25, and even the newer stuff I have that is not released.
And therefore is ahead of pretty much every engine not using a Stockfish network.
Only a handful of engines above it.

It definitely is ahead of Ethereal 13.25, as Koivisto 7.0 is around Ethereal 13.25 according to the rating lists, and Koivisto has gained at least 33 elo since 7.0 according to the regression tests on the Koivisto github wiki.

http://chess.grantnet.us/test/21734/
I would suspect that regression test is wildly over blown, due to the opening book and TC.
But I do agree anyway

SPCC shows Koivisto 7.5 to be 12 elo above Ethereal 13.25.

https://www.sp-cc.de/

[Jouni]

Longer time control (60/5) and new Koivisto:

Code: Select all

                                   
1   Fire 8.NN.MC.3 x64 avx2  +4/=32/-4 50.00%   20.0/40  400.00
2   Koivisto 7.9             +4/=32/-4 50.00%   20.0/40  400.00

WOW!

[Madeleine Birchfield]

Longer time control (60/5) and new Koivisto:

Code: Select all

                                   
1   Fire 8.NN.MC.3 x64 avx2  +4/=32/-4 50.00%   20.0/40  400.00
2   Koivisto 7.9             +4/=32/-4 50.00%   20.0/40  400.00

WOW!

Now you could run 360 more games to increase the statistical significance of your result.

[AndrewGrant]

Longer time control (60/5) and new Koivisto:

Code: Select all

                                   
1   Fire 8.NN.MC.3 x64 avx2  +4/=32/-4 50.00%   20.0/40  400.00
2   Koivisto 7.9             +4/=32/-4 50.00%   20.0/40  400.00

WOW!

I think SSS. Koivisto is near Ethereal, if not slightly stronger, and Ethereal cannot outplay Stockfish networks.
Unless you actually have MC mode enabled? In which case that would explain the gap being covered.

[matejst]

A question to Andrew:

I was always curious why you do not collaborate on a single project at -- to call it informally -- OpenBench? There are so many talented programmers, a lot of ideas -- and perhaps enough computing power, a must nowadays -- to seriously compete with the Lc0 and SF projects.

To be frank, I prefer several engines to only one -- but you already help each others, exchange ideas, and there will probably be some convergence, so why not?

[AndrewGrant]

A question to Andrew:

I was always curious why you do not collaborate on a single project at -- to call it informally -- OpenBench? There are so many talented programmers, a lot of ideas -- and perhaps enough computing power, a must nowadays -- to seriously compete with the Lc0 and SF projects.

To be frank, I prefer several engines to only one -- but you already help each others, exchange ideas, and there will probably be some convergence, so why not?

People like to do things themselves. It does not matter to me how much or how little Ethereal resembles Stockfish, because I know that I wrote every line of code myself one at a time, put thought into it, and tested it. I like building things. I built OpenBench from scratch even though I could have forked Fishtest, because I wanted to see the design decisions and make them myself. You don't appreciate the full scope of a project and its technical quirks when you are not there at the ground stage.

I know that others feel this way too. Yes, everyone on the framework could work on a single engine, and make it much stronger. I believe edging out Komodo is entirely doable in a short time-period -- I had almost done it myself prior to the advent of NNUE.

I would welcome everyone working on Ethereal. Just as Koivisto team would welcome everyone working on Koivisto, Just as Connor would welcome everyone working on Seer, I imagine.

Maybe its an ego thing?