TOGA IV with Rebel 14.2 NNUE evaluation

My experience for tournaments is also that a small book with lots of weird moves helps better then big book.
Big books are russian roulette.

Often it is enough to begin with e4 h6 or h3 instead of playing normal moves.
Then we can use kings gambit, blackmar diemer,  budapest opening, ….

There are gambit lines in almost any opening, and there are rare moves.

Target is to throw opponent program out of the book so that the interesting part begins.

Already made a start.

Mclane wrote:

My experience for tournaments is also that a small book with lots of weird moves helps better then big book.
Big books are russian roulette.

Often it is enough to begin with e4 h6 or h3 instead of playing normal moves.
Then we can use kings gambit, blackmar diemer, budapest opening, ….

There are gambit lines in almost any opening, and there are rare moves.

Target is to throw opponent program out of the book so that the interesting part begins.

I don’t think you have tried this in recent years with programs of similar strength at sth else but bullet time controls, Thorsten.
The Budapest and the Blackmar Diemer are just unsound by modern standards, and especially against the Budapest the usual engine moves work extremely well.
And 1. e4 h6? You just can’t do that.
I feel in a similar way about the kings gambit, but I don’t have enough experience with it to be sure.
1. h3 is probably kind of playable (at least I can’t think of a serious reason why it shouldn’t be), but this can’t be a good solution to the opening problem when you might as well play 1.c3 e.g. that is certainly more useful.
Engines have become really good at openings by now.
There are exceptions like the way Stockfish plays the Berlin with black on its own where probably any 2500 engine can get a draw against, so you would like to fix it with an opening book.
Anyway, Ed can certainly build a decent opening book as proven by his experiments. It’s just that no one who wants to win would like to have an opening book as engines did have in the 2000s where you got the ugly with the good.

for CSTAL it is important to have unbalanced opening situations.

whatever. i have at least 6 machines here without job waiting for some computation time.

Peter Berger wrote:

Mclane wrote:

My experience for tournaments is also that a small book with lots of weird moves helps better then big book.
Big books are russian roulette.

Often it is enough to begin with e4 h6 or h3 instead of playing normal moves.
Then we can use kings gambit, blackmar diemer,  budapest opening, ….

There are gambit lines in almost any opening, and there are rare moves.

Target is to throw opponent program out of the book so that the interesting part begins.

I don’t think you have tried this in recent years with programs of similar strength at sth else but bullet time controls, Thorsten.
The Budapest and the Blackmar Diemer are just unsound by modern standards, and especially against the Budapest the usual engine moves work extremely well.
And 1. e4 h6? You just can’t do that.
I feel in a similar way about the kings gambit, but I don’t have enough experience with it to be sure.
1. h3 is probably kind of playable (at least I can’t think of a serious reason why it shouldn’t be), but this can’t be a good solution to the opening problem when you might as well play 1.c3 e.g. that is certainly more useful.
Engines have become really good at openings by now.
There are exceptions like the way Stockfish plays the Berlin with black on its own where probably any 2500 engine can get a draw against, so you would like to fix it with an opening book.
Anyway, Ed can certainly build a decent opening book as proven by his experiments. It’s just that no one who wants to win would like to have an opening book as engines did have in the 2000s where you got the ugly with the good.

Both CSTAL and REBEL score 20-30 elo better using the gambit positions of the GRL than playing balanced positions as used for instance in CEGT and CCRL.

Admin wrote:

Both CSTAL and REBEL score 20-30 elo better using the gambit positions of the GRL than playing balanced positions as used for instance in CEGT and CCRL.

Interesting. I wonder if this is mainly due to evaluation or to search.

Peter Berger wrote:

Admin wrote:

Both CSTAL and REBEL score 20-30 elo better using the gambit positions of the GRL than playing balanced positions as used for instance in CEGT and CCRL.

Interesting. I wonder if this is mainly due to evaluation or to search.

Evaluation.

Admin wrote:

Peter Berger wrote:

Admin wrote:

Both CSTAL and REBEL score 20-30 elo better using the gambit positions of the GRL than playing balanced positions as used for instance in CEGT and CCRL.

Interesting. I wonder if this is mainly due to evaluation or to search.

Evaluation.

Both

Admin wrote:

Peter Berger wrote:

Admin wrote:

Both CSTAL and REBEL score 20-30 elo better using the gambit positions of the GRL than playing balanced positions as used for instance in CEGT and CCRL.

Interesting. I wonder if this is mainly due to evaluation or to search.

Evaluation.

This feels logical and you know much better than me anyway.
Just to explain the layman alternative to others: if you have a search tree like Stockfish’s ( or current Crafty’s just as I know its behaviour so well) the search tree is extremely shallow because of all the pruning. It can still grab the material (at least Stockfish can , and it can do a lot of more clever things, too). And if you win like a pawn in „usual“ positions, you win.
Maybe not so much in gambit positions.
I am currently playing Fritz 17 against Toga, they are behaving in very similar ways from a spectactor point of view. This results in a lot of draws (actually there is only one decided game so far). Fritz won it, but this was clearly just coincidence. The programs had their fail-lows and -highs at very similar times, my impression was that Fritz was just lucky that the position was in fact really just better for itself. Fritz 17 seems to be a very knowledgeable program btw though not the fastest, same as Toga/Rebel.

As far as i see it there is not yet a toga version with tal net, so only fruit with tal net that can only 1 core, right ?