personal experience regarding "self learning"

[deeds]

...
As usual, there are several ways to create an opening book, here are 2 of them:

- the fastest with the least human intervention (for tuning the book) is to use the tool expToBook

- the slowest (100% human intervention) but the safest is to use "SCID vs. PC".

1°) import PGN of all the learned openings


2°) open the tree window


3°) add moves

[deeds]

Which TC for learning / tournaments ?

Thanks to a learning with 7-8 threads, TC 1m+1s, 2000 games/opening, an engine can learn experience data that will resist up to tourneys at TC 4m+4s :

Code: Select all

   # PLAYER                             :  RATING  ERROR  POINTS  PLAYED   (%)    W    D    L  D(%)  OppAvg  OppN
   1 1m+1s Eman 8.20 nn-ad9b42354671    :       0   ----   302.5     600  50.4   81  443   76  73.8      -3     3
   2 4m+4s Stockfish 15.0 x64 BMI2      :       0     33   100.0     200  50.0   27  146   27  73.0       0     1
   3 3m+3s Stockfish 15.0 x64 BMI2      :      -4     34    99.0     200  49.5   25  148   27  74.0       0     1
   4 2m+2s Stockfish 15.0 x64 BMI2      :      -5     35    98.5     200  49.3   24  149   27  74.5       0     1

White advantage = -41.67 +/- 10.14
Draw rate (equal opponents) = 50.00 % +/- 0.00

tc odd - eman exp vs sf15


In fact, since NNUE era, the bestmoves of engines are already very strong even at shallow depths and the ELO gains are lower at greater depths :


elo gain vs depth+5 plies


For tourneys at longer TC, there are 2 solutions :

- easiest way, to reinforce your existing experience data

Code: Select all

   # PLAYER                             :  RATING  ERROR  POINTS  PLAYED   (%)    W    D    L  D(%)  OppAvg  OppN
   1 40T 1m+1s Eman 8.30 nn-ad9b42354671    :       0   ----   404.5     800  50.6  110  589  101  73.6      -4     4
   5 40T 5m+5s Stockfish 15.0 x64 BMI2      :      -7     34    98.0     200  49.0   25  146   29  73.0       0     1

White advantage = -40.94 +/- 9.05
Draw rate (equal opponents) = 50.00 % +/- 0.00

semi-monumental achievement

- longest way, to train the engines with a TC 2m+2s
(at the moment, i don't know until what TC such learned experience data can resist on tourneys...)

[deeds]

What does it mean for an engine to "learn an opening" ?

By learning, the engine will explore several lines of the opening.
In its experience file, it will add new moves, improve the statistics of moves it has already played, etc.
After a few dozen games (about 60 games for Eman), the main "key" positions specific to this opening will appear.
They will each contain several "close" moves.
The engine will have to play hundreds/thousands of games in order to test each combination several times.
The engine therefore learns an opening by finding the most effective combinations.
...

[deeds]

...
Here is an example with the "C11 French, Steinitz, BoleSlavesky variation" opening learned by Eman (LS7) :

[Event "Eman's learning"]
[Site "deeds"]
[Date "?"]
[Round "?"]
[White "eman"]
[Black "eman"]
[Result "*"]
[ECO "C11"]
[PlyCount "20"]

1. e4 e6 2. d4 d5 3. Nc3 Nf6 4. e5 Nfd7 5. f4 c5 6. Nf3 Nc6 7. Be3 Qb6 8.
Na4 Qa5+ 9. c3 b6 10. Bd2 c4 *

All engines played the same moves until this first "key" position :



Stockfish always played "b5" (b6b5) as its bestmove :




Eman played "0-0" (e8g8) as its bestmove, stored all its played moves after "0-0" (e8g8) on its experience file.
Then it explored the next "close" move => "b5" (b6b5) and stored all its played moves after "b5" (b6b5) on its experience file.


...

[deeds]

...
Obviously the "b5" line did not seem more effective so some will wonder why Eman did this.
The answer comes from the following "key" position :



Which is reached naturally by 14. ... 0-0 15. Nc3 b5.
But Eman also found an other combinaison to reach this "key" position.
After 14. ... b5 15. Nc3, where Stockfish mostly played "b4" (b5b4), Eman learned that "0-0" (e8g8) was slightly more effective :



Eman can therefore benefit from its experience after 14. ... 0-0 or 14. ... b5.
...

[deeds]

Evolution of the length of Eman's wins against Stockfish in this opening :



We can see how Stockfish loses faster and faster when Eman started to play with the 2 main lines (14. ... 0-0 and 14. ... b5).

[rulleeeee]

Where did you learn all this?

You are amazing

[deeds]

By practice ! Lol

When we try to understand, we learn faster.

Thank rulleeeee.

ps : Lately, I finally exiled myself from a community of saboteurs, cheats, censors. No need to read their false matchs, no need new engine's releases, no need to prove myself and i offered some of my own tools to the chess enthusiasts.

[rulleeeee]

Keep it up

[deeds]

How to get statistics of reinforce experience data ?

How to view/edit experience data ?

How to convert/export experience data ?

How to reinforce experience data ?

How do I make an opening book that matches an experience file ?