Some people download the client and run Leela locally to generate training games. These games are uploaded somewhere.
Leela NN processes the training games to improve itself. Every so often a new 'net' is produced.
Other people run different versions of the net ('the nets') against each other to measure progress. They also run the nets against other engines. These people are trying to determine just how strong Leela really is.
Somebody ('the developers') follows all of the above to improve different aspects of the Leela development chain.
Sometime soon:-
Someone has to decide what version of Leela will run in the next TCEC season.
I found relevant explanations of the process on various Leela support pages, specifically on the LeelaChessZero/lc0 Wiki:-
- (01) Home
- (02) FAQ
- (03) Neural Net Training
The numbers in parentheses map to the different steps of the process:-
(01) 'Getting-Started - Download and run Leela Chess Zero if you want to contribute or play games.'
(01) 'See recent self-play training games - Scroll to "Active Users", pick someone, then pick a game. These games are how Leela Chess Zero trains herself. They are played with extra randomness turned on so it can discover new good (and bad) moves. This means the quality of these games is lower than the match games.'
(03) 'The self-play games your client creates are used by the central server to improve the neural net. This process is called training (many people call the process of running the client to produce self-play games training, but in machine learning these games are only the input data for the actual training process). Some machine learning terms: [...]'
(01) 'See recent test match games - Click on the first row, first column, then pick a game. These games are played between recent versions of the engine to measure progress. They are blitz games played with 800 playouts (around 1 second) per move.'
(02) 'Current strength of Leela Chess Zero'
As for TCEC Season 12, at the time of last week's 'Finding Leela' post, the score in the final match was:-
+8-4=28 in favor of Stockfish
Since then another 40 games have been played, the score in these was:-
+14-3=23 idem, with ten Stockfish wins since the last Komodo win
Summing the scores over the two weeks gives a current score of +22-7=51 for Stockfish, i.e. 47.5 points of the 50.5 needed to win the match. This is with 20 games still to be played. With TCEC Season 13 due to start soon after the end of the Season 12 final, someone will have to make a decision about the Leela version even sooner than that.
No comments:
Post a Comment