Expositor Releases

[expositor]

Hello everyone!

Expositor 2BR17 has just been released; this is more or less the version that played in TCEC season 23 with some time control changes. Here are the release notes:

Version 2BR17 • Third Public Release

Stay safe, and thank you for your interest in Expositor!

Overview

This release features two major improvements:

• Search Tuning – Many of the constants in Expositor's search have now been tuned with a variation of simultaneous perturbation stochastic approximation (SPSA). Along with internal iterative reductions, this has increased the strength by approximately 100 Elo against 2WN29 in self-play.
• Internal Tablebases – Expositor now has an internal 3-man tablebase and I'm working on an optional 4-man tablebase. This will not increase her strength, since her endgame heuristics are sufficient for 3- and 4-man endgames (besides some KBN v K positions), but I wanted a break from work on HCE bootstrapping and thought it would be fun.

The complete list of changes made in this release is below.

Fixes

• fixed error where spurious moves could be printed at the end of the principal variation (PV)
• fixed error in Linux build script that caused the script to fail when building commits with commit dates in August through December

Improvements

• search constants have been tuned with a modified form of SPSA
• added internal iterative reductions
• transposition table entries are now prefetched
• added internal 3-man tablebase
• threads now clear their own context and statistics in parallel when a search begins
• the PGN parser can now handle moves written without a space after the move number (e.g. "1.e4")

Time Control

• target time adjustments based on PV stability now depend on the depth at which the PV changes
• searches are now less likely to stop immediately after the PV changes

Miscellaneous

• scores taper to half their nominal value rather than to zero as the 50-move horizon approaches
• made slight improvements to the efficiency of static exchange analysis
• made various small simplifications in the codebase
• added various developer tools

Let me know if there are any problems with the release or if you have any questions!

[Ajedrecista]

Hello Kade:

[...]

Let me know if there are any problems with the release or if you have any questions!

Congratulations for such a strong, new engine. Your version naming scheme caught my eye:

Code: Select all

Version	   Release Date
 2BR17     17 Sep 2022
 2WN29     29 May 2022
 2WQ23     23 Feb 2022

If you release on day D of month M of year Y, do you name the version like this?

Code: Select all

— 1st block [1 number]: the last digit of Y (Y mod 10).
— 2nd block [2 letters]:
  · 1st letter: 'W' for the first half-year (M =< 6), else 'B'. (Can be coded with an if statement).
  · 2nd letter: following the order {'K', 'Q', 'R', 'B', 'N', 'P'} inside the half-year. (Can be coded with modular arithmetic conditions in a nested if statement).
— 3rd block [2 numbers]: the day of release (D).

The second block would be like:

Code: Select all

01 → WK     07 → BK
02 → WQ     08 → BQ
03 → WR     09 → BR
04 → WB     10 → BB
05 → WN     11 → BN
06 → WP     12 → BP

Thank you in advance.

Regards from Spain.

Ajedrecista.

[expositor]

If you release on day D of month M of year Y, do you name the version like this?
...

Yep! That's exactly it ;)

[bastiball]

Hello everyone!

Expositor 2BR17 has just been released; this is more or less the version that played in TCEC season 23 with some time control changes. Here are the release notes:

Version 2BR17 • Third Public Release

Stay safe, and thank you for your interest in Expositor!

Overview

This release features two major improvements:

• Search Tuning – Many of the constants in Expositor's search have now been tuned with a variation of simultaneous perturbation stochastic approximation (SPSA). Along with internal iterative reductions, this has increased the strength by approximately 100 Elo against 2WN29 in self-play.
• Internal Tablebases – Expositor now has an internal 3-man tablebase and I'm working on an optional 4-man tablebase. This will not increase her strength, since her endgame heuristics are sufficient for 3- and 4-man endgames (besides some KBN v K positions), but I wanted a break from work on HCE bootstrapping and thought it would be fun.

The complete list of changes made in this release is below.

Fixes

• fixed error where spurious moves could be printed at the end of the principal variation (PV)
• fixed error in Linux build script that caused the script to fail when building commits with commit dates in August through December

Improvements

• search constants have been tuned with a modified form of SPSA
• added internal iterative reductions
• transposition table entries are now prefetched
• added internal 3-man tablebase
• threads now clear their own context and statistics in parallel when a search begins
• the PGN parser can now handle moves written without a space after the move number (e.g. "1.e4")

Time Control

• target time adjustments based on PV stability now depend on the depth at which the PV changes
• searches are now less likely to stop immediately after the PV changes

Miscellaneous

• scores taper to half their nominal value rather than to zero as the 50-move horizon approaches
• made slight improvements to the efficiency of static exchange analysis
• made various small simplifications in the codebase
• added various developer tools

Let me know if there are any problems with the release or if you have any questions!

Hello,
Any chances to implement chess960?

Regards,
Basti

[expositor]

Any chances to implement chess960?

Yes, I've thought about it! and it is on the to-do list, although I admit not very high priority at the moment.

I'm actually also a big fan of lkaufman's idea, Chess324, and I hope it catches on.

[bastiball]

Any chances to implement chess960?

Yes, I've thought about it! and it is on the to-do list, although I admit not very high priority at the moment.

I'm actually also a big fan of lkaufman's idea, Chess324, and I hope it catches on.

Yeah, I like the idea of Ikaufman's too about chess 324. Can't wait when you implement that, it is just I'm currently testing in our FRC list.