Train your engine !

[Eduard Nemeth]

Thanks for this interesting topic. What are your experiences with Brainlearn/Shashchess? What should be considered when training the Brainlearn learning file?

I am currently testing Brainlearn 20.1, but in a version I have modified.

A few weeks ago I created my own engine based entirely on Stockfish dev (current patches are Nov 23, 2022). Learning from Stockfish/MZ-Eman is integrated. I play with it on PlayChess.com.

I had created an empty learning file and played about 6000 live games 5+0 on the server so far.

Now I want to compare the Brainlearn learning file. For this I modified Brainlearn's search to be pretty much identical to my other engine. The only difference (I belive) is the learning file.

An empty learning file has also been created here, and it is played with the same settings (the parameters in Timeman are also identical).

[deeds]

What are your experiences with Brainlearn/Shashchess? What should be considered when training the Brainlearn learning file ?

Before playing on tourney, train 500 to 1000 games per opening. Use several sparring-partners (= different playing styles), if possible with learning feature in order to avoid that they play the same errors.

[deeds]

few Brainlearn's experiments

[deeds]

How can the engine's analysis be improved by its experience file ?

At the start of the game, the engine announces that such and such a move would be the best for the next 30 or 40 moves (D30, D40, etc.).
But he has no idea of its real effectiveness over the whole game.
He doesn't even know the length of the game or if his opponent will play the same moves as those it mentioned in his PV.

The experience file should help the engine to choose the most effective moves for the whole length of the game, even if these moves are not the best from the engine's analysis.

As we know that with the LazySMP search algorithm, we don't always get the same moves, we have to compare several PVs.
So here is 10 PVs analyzed by Eman 8.40, 8 threads, hash 4 GB, fixed D30.

From "1.g4" without experience file :

In red : moves that are not the most effective
white chances : after "1.g4" white normally loses so his percentage's chance should be as low as possible

From "1.g4" with an experience file :

In red : moves that are not the most effective
white chances : after "1.g4" white normally loses so his percentage's chance should be as low as possible

At first glance, with experience file, Eman announces less moves that are not the most effective and the chances of whites are less too.
Eman speed is -13% faster but its time-to-depth is faster (in average, it takes -4.2 sec to reach D30).

[deeds]

Here is a summary of the "learning session 8" openings :



You will find individual learning curves here.

learning session 8 :

Code: Select all

   # PLAYER                               :  RATING  ERROR   POINTS  PLAYED   (%)    W      D    L  D(%)  OppAvg  OppN
   1 Eman experience                      :       0   ----  33137.5   66000  50.2  366  65543   91  99.3      -1     4
   2 brainlearn 18.0 nn-ad9b42354671      :      -1      4   8237.0   16500  49.9   23  16428   49  99.6       0     1
   3 HypnoS 220822 nn-ad9b42354671        :      -1      4   8226.5   16500  49.9   29  16395   76  99.4       0     1
   4 Aurora 200822 nn-ad9b42354671        :      -1      4   8217.5   16500  49.8   26  16383   91  99.3       0     1
   5 ShashChess 23.1.1 nn-ad9b42354671    :      -3      4   8181.5   16500  49.6   13  16337  150  99.0       0     1

White advantage = 1.15 +/- 0.93
Draw rate (equal opponents) = 50.00 % +/- 0.00

Some words :

* In the previous 7 learning sessions, the TC was 1m+1s (avg. D27, 1806 ms/move, 192 sec/game). So to play 2000 games, it only took about 4-5 days/opening but I had a lot of data to reinforce (at D40).

* Since this 8th learning session, the TC is 2m+2s (avg. D36, 4236 ms/move, 348 sec/game). So to play 2000 games, it takes about 8-9 days/opening but I have much less data to reinforce (at D40).

* 2 new engines entered in the learning room :
- Aurora, courtesy of Ron DOUGHIE (Sarona), now replaces Stockfish as the reference engine so it started from scratch (empty experience file)
- Shashchess, by Andrea MANZO (amchess), now replaces JudaS (by Marco ZERBINATI) so it started from scratch (empty experience file)

* During this learning session, there are a lot less ELO gains between the engines because all the openings come from opening books or other experience files so they are particularly "drawish". I chose them to catch up with Eman because he had no experience in all these openings.

[ZamChess]

Thank you DeeDs, for your dedication and work for chess, which you share with us.

I wish you and your family all the best.

[deeds]

Thank you DeeDs, for your dedication and work for chess, which you share with us.

I wish you and your family all the best.

no thanks ! I haven't forgotten that you betrayed, you supported cheaters, you protected saboteurs :

http://outskirts.altervista.org/forum/v ... ost#p44874
http://outskirts.altervista.org/forum/v ... ost#p46109
http://outskirts.altervista.org/forum/v ... ost#p47506

We don't like you here too !

[ZamChess]

You have acted like an infatuated, you have apostroped everyone, you have triggered everything !
If you know/understand more about programming than others, it doesn’t mean you’re allowed to slander others who don’t understand.
And the fact that you don’t know how to forget, to let go of negative things, proves that you are not yet mature.

Conclusion: Very good about programming, very bad about character.

I hope you are healthy and I hope God will give you a better mind.

[deeds]

LS9 in progress...

learning farm #1 :


learning farm #2 :


A few more learning sessions and Eman's learning will be perfectly balanced (20% for each ECO group) :


And good luck to all those who train their engines without having removed the 2nd patch of November 5, 2022...

[rulleeeee]

What program do you teach Eman with?