What is the strongest neural-network-only chess engine? • page 1/1 • General Chess Discussion • lichess.org

In last years many chess engines use neural networks, however many of them use it as one of many components: stockfish, leela or alphazero use neural network for position evaluation or to select move to explore, but perform tree-searching with diffrent algorithms (as mcts); some of them (for example a0) can use ONLY neural network to make a move, but this makes it much weaker.

My question is: What is the strongest neural network that plays chess only by itself?

In last years many chess engines use neural networks, however many of them use it as one of many components: stockfish, leela or alphazero use neural network for position evaluation or to select move to explore, but perform tree-searching with diffrent algorithms (as mcts); some of them (for example a0) can use ONLY neural network to make a move, but this makes it much weaker. My question is: What is the strongest neural network that plays chess only by itself?

kalafiorczyk

#2

https://maiachess.com/

It uses Lc0 binary but configured to "go nodes 1", that is evaluating only single tree node.

https://maiachess.com/ It uses Lc0 binary but configured to "go nodes 1", that is evaluating only single tree node.

ContextManager

#3

@kalafiorczyk said in #2:

maiachess.com/

It uses Lc0 binary but configured to "go nodes 1", that is evaluating only single tree node.
Isn't maia supposed to play at 1100-1900 level? I'm almost sure that there are stronger networks.

@kalafiorczyk said in #2: > maiachess.com/ > > It uses Lc0 binary but configured to "go nodes 1", that is evaluating only single tree node. Isn't maia supposed to play at 1100-1900 level? I'm almost sure that there are stronger networks.

1100-1900?

@ContextManager said in #3:

Isn't maia supposed to play at 1100-1900 level? I'm almost sure that there are stronger networks.
What. Bro it is the code. If it played against let's say, 2500, it would play like tht

@ContextManager said in #3: > Isn't maia supposed to play at 1100-1900 level? I'm almost sure that there are stronger networks. What. Bro it is the code. If it played against let's say, 2500, it would play like tht

check-sandwich_1-2-3

#6

I think that the NNUE only came in at stockfish 12 so I maybe stockfish 11 could be an idea.

Then there's also the engines like kovisto, stoofvlees, houdini, RofChade and Ethereal but I am not sure if they are neural nets or just good at the game.

Hope this helps.

I think that the NNUE only came in at stockfish 12 so I maybe stockfish 11 could be an idea. Then there's also the engines like kovisto, stoofvlees, houdini, RofChade and Ethereal but I am not sure if they are neural nets or just good at the game. Hope this helps.

ContextManager

#7

@DERG_CHESS said in #5:

What. Bro it is the code. If it played against let's say, 2500, it would play like tht
Some time ago I tried to train it on lichess elite database (https://database.nikonoel.fr/), but it didn't seem to improve a lot compared to 1900 version (maybe I did something wrong).

@DERG_CHESS said in #5: > What. Bro it is the code. If it played against let's say, 2500, it would play like tht Some time ago I tried to train it on lichess elite database (https://database.nikonoel.fr/), but it didn't seem to improve a lot compared to 1900 version (maybe I did something wrong).