Orion 0.8 64-bit Gauntlet for CCRL 40/15

[Graham Banks]

PGN - http://kirill-kryukov.com/chess/discuss ... p?id=47766

[Graham Banks]

Code: Select all

CCRL 40/15 Rating List - Custom engine selection
1209412 games played by 2772 programs, run by 23 testers
Ponder off, General books (up to 12 moves), 3-4-5 piece EGTB
Time control: Equivalent to 40 moves in 15 minutes on an Intel i7-4770k.
Computed on January 24, 2021 with Bayeselo based on 1'209'412 games
Tested by CCRL team, 2005-2021, http://ccrl.chessdom.com/ccrl/4040/

Rank                 Engine                   Elo   +    -   Score  AvOp  Games
1 Orion 0.8 64-bit                        3126  +25  -25  49.8%   -2.0   479
  Orion 0.7 64-bit                        2761  +25  -24  51.0%   -4.2   526
  Orion 0.6 64-bit                        2629  +19  -19  49.1%   +5.7   915
  Orion 0.5 64-bit                        2527  +14  -14  46.6%  +24.4  1761
  Orion 0.4 64-bit                        2452  +20  -20  49.7%   +4.4   866
  Orion 0.3.0 64-bit                      2388  +19  -19  48.9%   +8.3   962
  Orion 0.2 64-bit                        2279  +21  -21  48.3%  +12.7   813

[David Carteau]

Thanks a lot Graham !

I'm really happy with these results, and with the performance of my Cerebrum library !

Since v0.8 has been released, I discovered a bug in my KPK endgame handcrafted code, which can lead to lose some games. But, that's... life !

My current efforts are entirely headed to learning without any external knowledge :

- first, trying a "zero" approach where the weights of the neural network are learnt through playing tons of games and selecting best networks (genetic approach) : this is really slow, and I wonder if - due to the high number of weights - this can lead somewhere (after almost 1.5 month, current "elite" is not even able to win against a pure material evaluation function...)

- but also trying a "reinforcement learning" approach, where I have first evaluated a few million positions with a depth 8 search and a pure material evalution function, and then train a neural network on the result. To my surprise, I almost reached with this method the strength of Orion 0.1 (~2150 elo). I'm currently running the 2nd round (evaluate a few million positions with the resulting network at depth 8, and then re-train another net, etc.)

I hope all these efforts will pay !