New J.A Compiles

Discussion of anything and everything relating to chess playing software and machines.

Moderator: Ras

User avatar
Jim Ablett
Posts: 2360
Joined: Fri Jul 14, 2006 7:56 am
Location: London, England
Full name: Jim Ablett

Re: New J.A Compiles

Post by Jim Ablett »

Werner wrote: Sun Nov 23, 2025 11:11 am
Jim Ablett wrote: Wed Nov 19, 2025 9:19 pm An update to my experimental nnue branch.
Add hybrid nnue mode.
Jim.
Hi Jim, is it possible, the net is not (correct) used? I got following results (same as version without a net):

1 Dirty Catpaw x64 +147 +53/=34/-13 70.00% 70.0/100 (2898)
2 Olithink 5.11.9 -147 +13/=34/-53 30.00% 30.0/100 performance 2751 +36

Olithink NNUEhybrid 2025 (use NNUE + Hybrid Eval)
1 Dirty Catpaw x64 +143 +50/=39/-11 69.50% 69.5/100
2 Olithink 5.11.9NNh x64 -143 +11/=39/-50 30.50% 30.5/100 2755

Olithink NNUE 2025 (default, only the net)
1 Dirty Catpaw x64 +135 +47/=43/-10 68.50% 68.5/100
2 Olithink 5.11.9NN x64 1CPU -135 +10/=43/-47 31.50% 31.5/100 2763
Hi Werner,

I'll check. Thanks for the test.

Jim.
User avatar
Jim Ablett
Posts: 2360
Joined: Fri Jul 14, 2006 7:56 am
Location: London, England
Full name: Jim Ablett

Re: New J.A Compiles

Post by Jim Ablett »

Image
https://phalanx.sourceforge.net/
Phalanx XXV UCI by Dusan Dobes

Phalanx XXV UCI 64 JA

Update:
Added UCI skill level system for human player.

Key Features
ELO Range: 1000 (beginner) to 2700 (maximum strength), with 18 calibrated profiles in between

Strength Adjustments:
Depth limiting: Weaker levels search shallower (1000 ELO searches to ~200 plies, 2700 searches to ~10000)
Evaluation noise: Random perturbations added to evaluations (-80 to +80 cp for beginners, 0 for strong levels)
Smooth scaling: Linear progression creates realistic strength transitions
-----------------------------------------------------------------

3-fold draw avoidance logic.
When a 3-fold repetition would occur, the engine calculates the current position evaluation
If ahead by 300+ centipawns (configurable), it refuses the draw and continues searching
If not ahead by the threshold, it accepts the draw normally

Configurable Threshold:
Default: 300 centipawns
GUI can adjust via: setoption name RepetitionThreshold value X
Range: 0-2000 centipawns
Value 0 disables the feature entirely

Strategic Benefits:
Engine won't accept draws when clearly winning
Maintains ability to accept draws when behind
Can be tuned per engine strength or opening book preference
Configurable by GUI users
---------------------------------------------------------------

Fixes to mate search.
Proton link:
https://drive.proton.me/urls/NPS4958HWW#DkVXbeNnkdg1

Smash link:
https://fromsmash.com/Phalanx-XXV64JA

Jim.
User avatar
Werner
Posts: 3004
Joined: Wed Mar 08, 2006 10:09 pm
Location: Germany
Full name: Werner Schüle

Re: New J.A Compiles

Post by Werner »

Olithink NN Old 2025 the first version from JA: olithink_5116_nnue_bmi2_ja.exe 472.064 14.09.2025 19:58 -a--
1 Olithink 5.11.6nnue old +70 +40/=40/-20 60.00% 60.0/100 (2967) +216 !!
2 Dirty Catpaw x64 -70 +20/=40/-40 40.00% 40.0/100 (2897)
Werner
User avatar
Jim Ablett
Posts: 2360
Joined: Fri Jul 14, 2006 7:56 am
Location: London, England
Full name: Jim Ablett

Re: New J.A Compiles

Post by Jim Ablett »

Werner wrote: Sun Nov 23, 2025 5:04 pm Olithink NN Old 2025 the first version from JA: olithink_5116_nnue_bmi2_ja.exe 472.064 14.09.2025 19:58 -a--
1 Olithink 5.11.6nnue old +70 +40/=40/-20 60.00% 60.0/100 (2967) +216 !!
2 Dirty Catpaw x64 -70 +20/=40/-40 40.00% 40.0/100 (2897)
I've found a bad bug that made engine always fall back to traditional eval. Just running some tests now.
Update released soon.

Thanks Werner,
Jim.
SlonoSlonar
Posts: 24
Joined: Sat Feb 01, 2025 10:23 am
Full name: Pavel Podlozniy

Re: New J.A Compiles

Post by SlonoSlonar »

Jim Ablett wrote: Sat Nov 22, 2025 12:25 pm Image
https://github.com/codedeliveryservice/Reckless
Reckless 0.9.0 Dev.967b by Arseniy Surkov

Reckless 0.9.0 Dev.967b 64 JA

Windows & Linux profiled builds

Proton link:
https://drive.proton.me/urls/XD3PS1YK7C#rWUbWOyRZxjW

Smash link:
https://fromsmash.com/Reckless090Dev967bJA

Jim.
Unfortunately, both the author's and your Reckless compilations return the "00000005" error when opening the engine in the console and are completely inoperable.

It was also mentioned above that it doesn't open in Fritz.
I'm writing this only to raise awareness of the issue, and in no way criticize you or the engine creator.
Owners of older hardware are facing increasing problems of this kind; progress is unstoppable.

This is especially frustrating here, as it's currently the second-most powerful game engine, slightly ahead of PlentyChess and Obsidian.
User avatar
Sylwy
Posts: 4916
Joined: Fri Apr 21, 2006 4:19 pm
Location: IAȘI - the historical capital of MOLDOVA
Full name: Silvian Rucsandescu

Re: New J.A Compiles

Post by Sylwy »

SlonoSlonar wrote: Sun Nov 23, 2025 10:49 pm Unfortunately, both the author's and your Reckless compilations return the "00000005" error when opening the engine in the console and are completely inoperable.
Is that so ?

Image

Image
User avatar
Jim Ablett
Posts: 2360
Joined: Fri Jul 14, 2006 7:56 am
Location: London, England
Full name: Jim Ablett

Re: New J.A Compiles

Post by Jim Ablett »

Sylwy wrote: Mon Nov 24, 2025 9:51 am
SlonoSlonar wrote: Sun Nov 23, 2025 10:49 pm Unfortunately, both the author's and your Reckless compilations return the "00000005" error when opening the engine in the console and are completely inoperable.
Is that so ?

Image

Image
Hi Sylwy,

He's using old hardware.
Because Reckless in written in Rust I can't use libraries like 'SIMDe Everywhere' to compile it to work on old cpus.

Jim.