zurichess jura released

[brtzsnr]

Hello, everyone!

I'm happy to release a new version of my engine, zurichess jura.

The source code for this release can be browsed from https://bitbucket.org/zurichess/zuriche ... lease.jura
The release binaries can be downloaded at: https://bitbucket.org/zurichess/zurichess/downloads
Linux AMD64 https://bitbucket.org/zurichess/zuriche ... inux-amd64
Windows AMD64 https://bitbucket.org/zurichess/zuriche ... -amd64.exe


An overview of the most important changes is:

• * Improve search:
  • * Implement razoring at depth 1.
    * Implement counter move heuristics.
    * Tweak LMR, NMP and FP.
  * Improve evaluation:
  • * Retune using training data at https://bitbucket.org/zurichess/tuner/downloads.
    * Add King-Queen tropism.
    * Add Rook PSqT.
  * Evaluate connected pawns based on their file.
  * Tweak time control to reduce more of the forfeits.
  * Usual code clean ups, speed ups and bug fixes.
  * New Library API: added Position.GivesCheck

In self play at fast controls it's about 85 Elo stronger. Self-play strength improvement
does not necessary indicate an improvement against other engines.

Code: Select all

700 @ 40/15+0.15
316 - 148 - 236
ELO 85.04±21.29

Despite the Elo improvement, the code has not grown too much.

Code: Select all

 27 files changed, 872 insertions(+), 660 deletions(-)

Please test the new version if you maintain a rating list.

Many thanks to everyone who included zurichess in their private tests and tournaments and made this release possible.

Regards,

[brtzsnr]

Regarding future releases:

1. I will switch to release when the improvement gets closer to 50 Elo, instead of 100 Elo. The engine is a lot stronger than the first Aargau release and improvements are harder to find and often very small.
2. SMP will come after the engine reaches 3000+ Elo on CCRL 40/40 in single thread mode. I believe evaluation and search can still be improved.
3. Lately I've been exploring automatic feature extraction for evaluation. Of course, there is very little literature on this topic so no advances so far. Long term I want to move away from manual feature engineering.

[Werner]

Thanks Alexandru,

I will start testing soon...

[SzG]

Thanks Alexandru.

Just to be sure, is the default skill level 0 the strongest?

[styx]

I asked the same question some weeks ago: yes it is

It's maybe an unfavourable decision to name it "skill level" (as in stockfish). Maybe "handicap level" would cause less confusion.

[SzG]

I asked the same question some weeks ago: yes it is

It's maybe an unfavourable decision to name it "skill level" (as in stockfish). Maybe "handicap level" would cause less confusion.

Thanks Andreas.

[Graham Banks]

Thanks Alexandru.

[BrendanJNorman]

Many thanks to everyone who included zurichess in their private tests and tournaments and made this release possible.

Regards,

Many thanks to you for your interesting engine.

I like it a lot (a unique human-like playing style!), so I made a logo for you in thanks. I hope you like it.

http://chessncognac.com/wp-content/uplo ... ss-128.jpg

[brtzsnr]

Brendan~

Thank you for your kinds words. I read your blog from time to time, I didn't know you are also a fan of computer chess. I haven't trained the engine for any particular style so it's just a coincidence. Would you mind if I put your praise on the engine's website?

[brtzsnr]

That's a good idea. I might rename the option to avoid further confusion. I don't like non-zero defaults as a programming principle which is why Skill level defaults to 0.