I wonder if there is a rating list for learning chess engines

[Uri Blass]

The idea is not to test the rating of the engine but the rating of the engine plus learning when engines are allowed to learn.

In order to maximize the effect of learning I suggest that for every position every 2 engines play 20 games from the same position with no book when you count only 2 games for rating.

For example if I want to test A against B after 1.e4 c5 with no book then A play against B 10 times with white when only the last game is counted for rating and A play against B 10 times with black when only the last game is counted for rating.

I wonder what are the engines that can earn rating from this type of tournament except RomiChess.

It is possible to use also Stockfish that has no learning function in the rating list and it may be interesting to see if some engine can beat it thanks to learning or not(or maybe 10 games are not enough to beat stockfish with no learning and you usually need 100 games or 1000 games even with strong engines and a good learning function).

For time control I suggest to start with some fast time control and if something work then try it also with a slower time control.

[Jouni]

NNUE is the best learning method !

[Mike Sherwin]

The idea is not to test the rating of the engine but the rating of the engine plus learning when engines are allowed to learn.

In order to maximize the effect of learning I suggest that for every position every 2 engines play 20 games from the same position with no book when you count only 2 games for rating.

For example if I want to test A against B after 1.e4 c5 with no book then A play against B 10 times with white when only the last game is counted for rating and A play against B 10 times with black when only the last game is counted for rating.

I wonder what are the engines that can earn rating from this type of tournament except RomiChess.

It is possible to use also Stockfish that has no learning function in the rating list and it may be interesting to see if some engine can beat it thanks to learning or not(or maybe 10 games are not enough to beat stockfish with no learning and you usually need 100 games or 1000 games even with strong engines and a good learning function).

For time control I suggest to start with some fast time control and if something work then try it also with a slower time control.

Stockfish even only using one thread is non deterministic. I don't know why. The best RomiChess has been able to get is a draw or two after several thousand games. Isn't SF vs SF mostly draws? I think RomiChess can still learn to at least draw all games but RomiChess would need to train for millions of games. More interesting would be to train for millions of games against Leela and then play other engines.

[Mike Sherwin]

NNUE is the best learning method !

Right now it is but it is nowhere near optimal. Only real time reinforcement learning can approach optimal. A competent implementation of real time learning will one day be shown to add at least another 1000 elo.