deeds_chess_system
Re: SF15.1 at TC 60+1 vs Eman exp. at fixed depth
The rest of this experiment can be found here
Re: deeds_chess_system
hardware : dual xeon e5-2660v3
software : cutechess 1.2.0
settings : TC 60s+1s, 40 threads, hash 4096 MB, partial syzygy 7men
Opening list : Jeroen Noomen's 2017 selection 100 openings
link : games
only for friends and family (request password and information by PM)
software : cutechess 1.2.0
settings : TC 60s+1s, 40 threads, hash 4096 MB, partial syzygy 7men
Opening list : Jeroen Noomen's 2017 selection 100 openings
Code: Select all
# PLAYER : RATING ERROR POINTS PLAYED (%) W D L D(%) OppAvg OppN
1 deeds_chess_system : 0 ---- 106.5 200 53.3 15 183 2 91.5 -23 1
2 stockfish 15.1 : -23 36 93.5 200 46.8 2 183 15 91.5 0 1
White advantage = 26.45 +/- 17.34
Draw rate (equal opponents) = 50.00 % +/- 0.00
only for friends and family (request password and information by PM)
Re: deeds_chess_system
After 8 opening's lists :
Now that the D.C.S. "proof-of-concept" phase is over, It's time to do a partial lifting on the NDA of the DeeDs Chess System :
*** Prerequisite knowledge ***
TRAINING MODE (=default settings)
- By default, engines with learning features are set to regularly try new moves.
- Unlike all amnesiac engines like Stockfish, LC0, Komodo Dragon, etc, the learning engines don't always play the moves with the best evaluated scores.
- Sometimes the new moves are more effective, sometimes they are worse but in all cases they contribute to update the experience data.
As it is unlikely that tournaments of daily testers will allow them to play 500 times each opening (=minimum required to learn an opening), without knowing their effectiveness, each move is risky. Whenever you see tournaments (e.g. : link #1, link #2, link #3) with learning engines set to their default settings, the tournament's result will represent more the engine's evaluation than the quality of its experience.
TOURNAMENT MODE (=experience file as book, deeds chess system, etc.)
- when the positions have experience data, the learning engines must play their most effective moves and above all not risk trying others.
- when the positions do not have experience data, the learning engines evaluate and then add new moves to their experience file.
Code: Select all
# PLAYER : RATING ERROR POINTS PLAYED (%) W D L D(%) OppAvg OppN
1 deeds_chess_system : 0 ---- 616.5 1080 57.1 163 907 10 84.0 -51 1
2 stockfish 15.1 : -51 14 463.5 1080 42.9 10 907 163 84.0 0 1
White advantage = 44.30 +/- 7.64
Draw rate (equal opponents) = 50.00 % +/- 0.00
Code: Select all
002 players : avg. D34, 001 795 ms/move, 001 080 games, 00 107 088 moves
deeds_chess_system : avg. D33, 001 989 ms/move, 001 080 games, 00 052 826 moves
stockfish 15.1 : avg. D34, 001 606 ms/move, 001 080 games, 00 054 262 moves
Averages :
0 178 sec/game (3 min/game)
117 plies/game (099 played plies/game)
Game stages :
- opening @ avg. D28, 002 263 ms/move
- middleg @ avg. D31, 001 782 ms/move
- endgame @ avg. D41, 001 375 ms/move
Now that the D.C.S. "proof-of-concept" phase is over, It's time to do a partial lifting on the NDA of the DeeDs Chess System :
*** Prerequisite knowledge ***
TRAINING MODE (=default settings)
- By default, engines with learning features are set to regularly try new moves.
- Unlike all amnesiac engines like Stockfish, LC0, Komodo Dragon, etc, the learning engines don't always play the moves with the best evaluated scores.
- Sometimes the new moves are more effective, sometimes they are worse but in all cases they contribute to update the experience data.
As it is unlikely that tournaments of daily testers will allow them to play 500 times each opening (=minimum required to learn an opening), without knowing their effectiveness, each move is risky. Whenever you see tournaments (e.g. : link #1, link #2, link #3) with learning engines set to their default settings, the tournament's result will represent more the engine's evaluation than the quality of its experience.
TOURNAMENT MODE (=experience file as book, deeds chess system, etc.)
- when the positions have experience data, the learning engines must play their most effective moves and above all not risk trying others.
- when the positions do not have experience data, the learning engines evaluate and then add new moves to their experience file.
Re: deeds_chess_system
*** Experience file as book ***
For those who want an overview of the quality of their engine's experience, you can activate the "Experience book" experimental feature.
Then, you have to configure some options :
Experience Book Min Depth
=> set a shallow depth to allow your engine to play silly moves (like do some unenlightened testers)
=> set a great depth to force your engine to respect the reinforced experience data
Experience Book Eval Importance
=> set a low value if you believe in the efficiency of your experience data
=> set a high value if you believe in the evaluation of your engine
Experience Book Max Moves
=> set a low value if you have trained your engine little (<500 games/opening)
=> set a high value if you have trained your engine a lot (>1000 games/opening)
For those who want an overview of the quality of their engine's experience, you can activate the "Experience book" experimental feature.
Then, you have to configure some options :
Experience Book Min Depth
=> set a shallow depth to allow your engine to play silly moves (like do some unenlightened testers)
=> set a great depth to force your engine to respect the reinforced experience data
Experience Book Eval Importance
=> set a low value if you believe in the efficiency of your experience data
=> set a high value if you believe in the evaluation of your engine
Experience Book Max Moves
=> set a low value if you have trained your engine little (<500 games/opening)
=> set a high value if you have trained your engine a lot (>1000 games/opening)
Re: deeds_chess_system
*** DeeDs Chess System ***
For all the positions of a game, the evaluation of the engine does not have the same precision.
The same goes for experience data, sometimes there are many, sometimes there are few.
Sometimes they are reinforced with refined scores at very great depths.
So with values fixed for all the positions of a game like those of the "Experience Book Min Depth / Eval Importance" options, it is not optimal.
Engine's moves are not always in conformity with its experience data.
D.C.S has been designed to make the best use of experience data in each position.
As i don't read the C/C++ language and i don't know anyone who can modify the source codes of the private engines,
i coded a tool which drives several engines simultaneous, which chooses the most efficient moves or the moves of the reinforced data or the moves from the opening book.
The move's choice depends on the position of the game, the among/quality of experience data, what contains the opening book, etc.
D.C.S passes UCI commands regarding opening book options to the "book_engine" configured to choose moves from the opening book.
D.C.S passes other UCI commands regarding engines settings to the "playing_engine" configured to analyze the positions, choose moves from the experience file, save new moves and update the experience data, etc.
D.C.S needs an INI file which looks like this :
Throughout this "proof-of-concept" phase, i used Eman 8.40 as playing_engine, asmFish 291118 as book_engine and Depth4_180423.exp as the experience file.
The rest of the D.C.S. features/details remains under NDA...
For all the positions of a game, the evaluation of the engine does not have the same precision.
The same goes for experience data, sometimes there are many, sometimes there are few.
Sometimes they are reinforced with refined scores at very great depths.
So with values fixed for all the positions of a game like those of the "Experience Book Min Depth / Eval Importance" options, it is not optimal.
Engine's moves are not always in conformity with its experience data.
D.C.S has been designed to make the best use of experience data in each position.
As i don't read the C/C++ language and i don't know anyone who can modify the source codes of the private engines,
i coded a tool which drives several engines simultaneous, which chooses the most efficient moves or the moves of the reinforced data or the moves from the opening book.
The move's choice depends on the position of the game, the among/quality of experience data, what contains the opening book, etc.
D.C.S passes UCI commands regarding opening book options to the "book_engine" configured to choose moves from the opening book.
D.C.S passes other UCI commands regarding engines settings to the "playing_engine" configured to analyze the positions, choose moves from the experience file, save new moves and update the experience data, etc.
D.C.S needs an INI file which looks like this :
Throughout this "proof-of-concept" phase, i used Eman 8.40 as playing_engine, asmFish 291118 as book_engine and Depth4_180423.exp as the experience file.
The rest of the D.C.S. features/details remains under NDA...
Re: opening book vs experience file
hardware : dual xeon e5-2660v3
software : cutechess-gui 1.2.0
settings : TC 60s+1s, 40 threads, hash 4096 MB, partial syzygy 7men
Opening list : startpos
link : games
pass : 59Ef9tOGT1RCBGYgpMTuzw
software : cutechess-gui 1.2.0
settings : TC 60s+1s, 40 threads, hash 4096 MB, partial syzygy 7men
Opening list : startpos
Code: Select all
# PLAYER : RATING ERROR POINTS PLAYED (%) W D L D(%) OppAvg OppN
1 Depth4_190623.exp (Eman 8.40) : 0 49 50.0 100 50.0 0 100 0 100.0 0 1
2 hcanfree-upd.bin (Eman 8.40) : 0 ---- 50.0 100 50.0 0 100 0 100.0 0 1
White advantage = 0.00 +/- 25.05
Draw rate (equal opponents) = 50.00 % +/- 0.00
pass : 59Ef9tOGT1RCBGYgpMTuzw
-
Sedat Canbaz
- Posts: 2407
- Joined: Wed Jun 21, 2023 6:29 am
Re: opening book vs experience file
Sorry to say that,deeds wrote: Tue Jun 20, 2023 8:55 pm hardware : dual xeon e5-2660v3
software : cutechess-gui 1.2.0
settings : TC 60s+1s, 40 threads, hash 4096 MB, partial syzygy 7men
Opening list : startposlink : gamesCode: Select all
# PLAYER : RATING ERROR POINTS PLAYED (%) W D L D(%) OppAvg OppN 1 Depth4_190623.exp (Eman 8.40) : 0 49 50.0 100 50.0 0 100 0 100.0 0 1 2 hcanfree-upd.bin (Eman 8.40) : 0 ---- 50.0 100 50.0 0 100 0 100.0 0 1 White advantage = 0.00 +/- 25.05 Draw rate (equal opponents) = 50.00 % +/- 0.00
pass : 59Ef9tOGT1RCBGYgpMTuzw
On what kind of planet do you live ? if not so clear:
You are running via completely unfair conditions!
Private vs Public.. hehe ))
Come on SCCT conditions...and let's see
Under real Arena.. what is going on...!
Re: opening book vs experience file
And it was only a little "warmup" with Eman only set by default (=training mode).
I didn't even need to use the experience file as a book so Depth4_190623.exp often had less time to play.
But hey, I'm nice, the games are available...
For engine trainers :
Yes, you read well !
From the start position (=no opening moves or others), with only 80% of conformity with the experience data of Depth4_190623.exp, it was enough to draw it.
I didn't even need to use the experience file as a book so Depth4_190623.exp often had less time to play.
But hey, I'm nice, the games are available...
For engine trainers :
Yes, you read well !
From the start position (=no opening moves or others), with only 80% of conformity with the experience data of Depth4_190623.exp, it was enough to draw it.