asean chess engines and rating list

[gbtami]

PyChess plays ASEAN/Makruk/Cambodian/Sittuyin variants also.

This reminds me: I adjusted some of the Sittuyin rules in Sjaak based on user input. There are also official tournament rules floating around.
I think I mentioned the changes here in a post, but I can also send you the official rules document, if you're interested.

Unfortunately I had no time to adjust PyChess but filed it on https://github.com/pychess/pychess/issues/1396
Thx. again!

[belzedar94]

I just ran a quick tournament to benchmark the available engines for this variant.

Match Fairy-SF 11.2 (23/09/2020) vs latest Fairy-SF dev vs SjaakII 1.4.1 vs Silangan v0-14
Variant: Asean Chess
Tournament: Round-Robin
Binaries: Fairy-SF 11.2, Fairy-SF dev 070121, SjaakII 1.4.1, Silangan v0-14
Engine Options: 1 thread, 2048 hash, no ponder
Hardware: i7-9750H
Time control: 30+0,3s
Book: Asean.epd
Tablebases: None
Adjudication: Resign if >1500cp for more than 5 moves
Software: Cutechess

Results:

Code: Select all

Rank Name                                                  Elo     +/-   Games    Wins  Losses   Draws   Points   Score    Draw   White   Black 
   1 Fairy-SF 11.2 Release (230920) - Classical Eval       306      43     300     225      13      62    256.0   85.3%   20.7%   86.0%   84.7% 
   2 Fairy-SF 070121 - Classical Eval                      250      41     300     211      26      63    242.5   80.8%   21.0%   81.7%   80.0% 
   3 Silangan v0-14                                       -118      41     300      99     197       4    101.0   33.7%    1.3%   33.3%   34.0% 
   4 SjaakII (check variant)                             -1111     nan     300       0     299       1      0.5    0.2%    0.3%    0.0%    0.3% 

600 of 600 games finished.

ORDO:

Code: Select all

   # PLAYER                                             :  RATING  POINTS  PLAYED   (%)
   1 Fairy-SF 11.2 Release (230920) - Classical Eval    :   882.3   256.0     300    85
   2 Fairy-SF 070121 - Classical Eval                   :   835.4   242.5     300    81
   3 Silangan v0-14                                     :     0.0   101.0     300    34
   4 SjaakII (check variant)                            :  -930.5     0.5     300     0

White advantage = 9.91
Draw rate (equal opponents) = 50.00 %

Further details and other variants testing:
https://docs.google.com/spreadsheets/d/ ... sp=sharing

Link to the project's discord (Fairy-Stockfish) - In case you are curious about chess variants or want to contribute:
https://discord.gg/fmCWv8PG

[IanO]

Are there any Leela-based engines for any of the southeast-asian chess variants? Since these are all more strategic, less tactical, I would think NN would be a good fit. I'd be curious if such engines would discover the same opening strategies used in Makruk, or play the endgames better than in chess due to the more stringent material counting rules in Makruk.

[belzedar94]

Are there any Leela-based engines for any of the southeast-asian chess variants? Since these are all more strategic, less tactical, I would think NN would be a good fit. I'd be curious if such engines would discover the same opening strategies used in Makruk, or play the endgames better than in chess due to the more stringent material counting rules in Makruk.

Hello Iano,

Indeed we tried to train some nets in Makruk and Makpong so far. The Makpong one was a success! After 4 versions, it has gained 300-400 elo over classical Stockfish evaluation. You can see the detail of the trained nets in the below link.

https://docs.google.com/spreadsheets/d/ ... sp=sharing

For Makruk success was not that high though. The drawish nature of the game and the already strong classical evaluation from Fairy-Stockfish make no room for improvement to the Neural Networks. Still, Fairy-Stockfish classical evaluation is already 400 elo above any other existing engine, so it plays at a godlike level as you can see in the link below:

https://github.com/ianfab/Fairy-Stockfi ... g-strength

If you would like to propose any other variant to be tested (Training a net for it) or contribute in any possible way, I'll share the link to the discord server of the Fairy-Stockfish project. You can find detailed information in the #Resources channel.

https://discord.gg/EpysNCe4

[IanO]

Fairy-Stockfish

That is NNUE, not the LeelaZero NN + MCTS I was inquiring about. Still, good to know that classical search techniques + NNUE work well on these games. (First time I'd heard about the Makpong variant.)