End of Era is there: SF is finally beaten!

[Alexander Schmidt]

Leela learns patterns in self play games. In new positions she tries to find this patterns. That's why she is running faster with gpu which are optimized for calculating images. Leela don't save one single position in it's NN. She found something in the position which made her believe is winning. We will likely never know what it was.

NN of Leela does not consist of "policy head" only...

It's just a very basic description since some fail to understand what a NN is.

[George Tsavdaris]

There is no any "spectacular" and there is no any "plan" on the side of Leela.
In reality the position after 20.-c4 and the good continuation is in the 32930net of Leela.
To prove this I used the weak GTX 1060 to find the 21.Bg6 move.
From net30100 to net32000 Leela want to move 21.Be4, 21.Bxc4, 21.0-0.
But using net32500 Leela finds 21.Bg6 immediately!
So anywhere between net32000 and net32500 Leela has analysed yet this position during self play
and 21.Bg6 is in the NN of Leela.
Obviously Stockfish also would be much more stronger if it could use a good book instead of analysing every each position.

You have ZERO idea about how Neural Net engines work.

Does anybody? Can you make a non-tautological explanation in, say 300 words, “how neural net engines work” that we can cut out and put in a frame on the wall?

For one, the 75 million games with the resulting positions Leela saw in her training for test30, are ZERO compared to the 10^40 or so Chess positions that exist so the above stupid argument that Leela has seen this exact position before so it's like having an opening book, so she is kinda cheating is just 100% wrong.

[M ANSARI]

There is no any "spectacular" and there is no any "plan" on the side of Leela.
In reality the position after 20.-c4 and the good continuation is in the 32930net of Leela.
To prove this I used the weak GTX 1060 to find the 21.Bg6 move.
From net30100 to net32000 Leela want to move 21.Be4, 21.Bxc4, 21.0-0.
But using net32500 Leela finds 21.Bg6 immediately!
So anywhere between net32000 and net32500 Leela has analysed yet this position during self play
and 21.Bg6 is in the NN of Leela.
Obviously Stockfish also would be much more stronger if it could use a good book instead of analysing every each position.

You have ZERO idea about how Neural Net engines work.

Does anybody? Can you make a non-tautological explanation in, say 300 words, “how neural net engines work” that we can cut out and put in a frame on the wall?

For one, the 75 million games with the resulting positions Leela saw in her training for test30, are ZERO compared to the 10^40 or so Chess positions that exist so the above stupid argument that Leela has seen this exact position before so it's like having an opening book, so she is kinda cheating is just 100% wrong.

Yes very true! The weight file or "learning file" of Lc0 is around 45mb ... this is a tiny file when you think about all the possible moves there are until Lc0 produced the very nice Bg6! move. The idea that Lc0 is just a large file of memorized book openings is beyond silly as the file would have to be 1,000,000 x or even more x larger. Even then it would be on its own once out of book and would lose very rapidly. I think a better way to describe Lc0 is that in its learning phase of playing many games against itself, it discovers some patterns and some motifs that work better than others. In a case of an advanced pawn it would probably notice that if the pawn can be protected and the other side can't dislodge it, it causes enough damage to the other side to give a winning position. In the case of the game where it played Bg6!, I think it noticed that it had a pattern where the rook couldn't move for many moves and that because of that it could get an advantage and thus played it. Whatever it is that Lc0 is doing has some revolutionary and ground breaking effect on the game of chess. Of course there are still many flaws and weaknesses as well, but hey ... this thing is only like 10 months old. One year from now several things will change in Lc0 favor ... a lot of talented people will start looking more seriously at NN ... and the hardware available is going to explode in strength as AI starts going mainstream. Who is to say that Google had the best implementation of NN? If history is a guide, there will be better, faster and more efficient methods that will soon come around. The combination of improved software and improved hardware will really make NN engines untouchable. I can't imagine that the top 3 engines will just let this all pass by. If it works (and it does work) there is nothing that will stop Houdini or SF or Komodo from also implementing some sort of NN for parts of the games where it is helpful.

[glennsamuel32]

Yes very true! The weight file or "learning file" of Lc0 is around 45mb ... this is a tiny file when you think about all the possible moves there are until Lc0 produced the very nice Bg6! move. The idea that Lc0 is just a large file of memorized book openings is beyond silly as the file would have to be 1,000,000 x or even more x larger.

40948 is 52.8mb and contains data on 34,201,430 positions.
As a comparison, a ctg book of 2.27gb contains 35,464,353 positions.

While leela is a very interesting project, you just cannot compare it to AB engines, that use manual tuning for all positions.
Whether the network contains images or samples or patterns or some positional data, it is still a database (and a considerable one too)
Like another user pointed out, "How are we sure we are playing against chess engine and not the games database? And when does the multimillion game database end and the chess engine begin ?"

I instead enjoy leela's existence along with all the other AB engines, without trying to match them.

Let me ask you this...
If SF had a learn file containing depth and eval only, without any results, of 35 million positions, and an evaluation engine to select the best move, would that be considered as acceptable, in all these ongoing tournaments ?

[George Tsavdaris]

While leela is a very interesting project, you just cannot compare it to AB engines, that use manual tuning for all positions.
Whether the network contains images or samples or patterns or some positional data, it is still a database (and a considerable one too)

And why the AB that contains handcrafted positional data is not a database?
By your definition it IS also, so no problem....

If SF had a learn file containing depth and eval only, without any results, of 35 million positions, and an evaluation engine to select the best move, would that be considered as acceptable, in all these ongoing tournaments ?

As long as it does not contain an opening tree with specific "1-1" position-eval data for playing the opening then yes why not?

[glennsamuel32]

And why the AB that contains handcrafted positional data is not a database?
By your definition it IS also, so no problem....

You call software coding a database ?
Well done

[Leo]

Yes very true! The weight file or "learning file" of Lc0 is around 45mb ... this is a tiny file when you think about all the possible moves there are until Lc0 produced the very nice Bg6! move. The idea that Lc0 is just a large file of memorized book openings is beyond silly as the file would have to be 1,000,000 x or even more x larger.

40948 is 52.8mb and contains data on 34,201,430 positions.
As a comparison, a ctg book of 2.27gb contains 35,464,353 positions.

While leela is a very interesting project, you just cannot compare it to AB engines, that use manual tuning for all positions.
Whether the network contains images or samples or patterns or some positional data, it is still a database (and a considerable one too)
Like another user pointed out, "How are we sure we are playing against chess engine and not the games database? And when does the multimillion game database end and the chess engine begin ?"

I instead enjoy leela's existence along with all the other AB engines, without trying to match them.

Let me ask you this...
If SF had a learn file containing depth and eval only, without any results, of 35 million positions, and an evaluation engine to select the best move, would that be considered as acceptable, in all these ongoing tournaments ?

Glen, you are a genius. You have articulated what I have long thought.

[corres]

Only some note to your dialog:
1, Nobody wrote about Chessbase or other classical type of chess books. The dimension of a chess book
depend on the technique of its coding. A pure digital book is far smaller than a commercial .dtg book.
2, As in the games played by LC0 without outer opening book proved there a lot of similar game in the self
play of Leela also. So the the self play games need not so large place what you suppose.
3, Really, the most larger any kind of book can not contain the whole possible chess position. But even a book
with holes in which there are positions together with winning chance and move vector to continue with the
possibility of moves can give great help to search in finding good moves.
4, If you know how would deal the similarity problems in NN technique, please write down. A lot of peoples
like to know your idea, I think.

Sorry, but I should like to read the down write of Leela from her developer than to read your obstinacy.

[syzygy]

Like another user pointed out, "How are we sure we are playing against chess engine and not the games database? And when does the multimillion game database end and the chess engine begin ?"

The answer is: by understanding what Leela does. (Or, if one does not have the time or inclination to study how the NN approach functions, by accepting what those that do understand it tell them.)

If SF had a learn file containing depth and eval only, without any results, of 35 million positions, and an evaluation engine to select the best move, would that be considered as acceptable, in all these ongoing tournaments ?

Leela has exactly as much a "learn file" on 35 million positions as SF has a hand-tuned "learn file" on the billion or so games that have been played on fishtest.

[George Tsavdaris]

And why the AB that contains handcrafted positional data is not a database?
By your definition it IS also, so no problem....

You call software coding a database ?
Well done

You said "or some positional data" and coding e.g Rook on open file worth +0.02 more is positional data.
Eval function of AB's contains positional data. So it's no different in that aspect to the weights of the NNs.