New Little Goliath 3.15 with NN released

[mclane]

How do you know

[Gabor Szots]

How do you know

Played a short blitz match against Marvin.

[Jouni]

After starting exe I get:

info Neural Network ( GoliathNet0001.bin ) OK!
info Neural Network ( pruning data ) inactive!
info Neural Network ( extension data ) inactive!
info Neural Network ( pattern recognition data ) inactive!

But speed with 1 core is stunning 5Mnps in starting position - really!?

[Guenther]

BTW for people, who want to do some tests from cmd, you need to send 'isready' otherwise it will close immediately.

I am surprised it is compiled w/o popcount and no instructions after at least sse4.1.
Good for me though ;-)

[BrendanJNorman]

But speed with 1 core is stunning 5Mnps in starting position - really!?

Goliath has always been a REALLY fast engine. Even since the ancient versions back in 2000 or so. I used to like this engine a lot when I was a kid - still do now in my late 30s!

The new version with NN is quite positional and plays in a very unique style, but not overwhelmingly strong that it looks "alien".

Really impressed.

[Jouni]

You can't argue progress, some 1m+1s games :

Code: Select all

Score of LGEvolution_NN vs LGEvolution: 52 - 0 - 1 [0.991]
...      LGEvolution_NN playing White: 27 - 0 - 0  [1.000] 27
...      LGEvolution_NN playing Black: 25 - 0 - 1  [0.981] 26
...      White vs Black: 27 - 25 - 1  [0.519] 53
Elo difference: 808.5 +/- nan, LOS: 100.0 %, DrawRatio: 1.9 %
58 of 200 games finished.

NN version has a higher node speed actually!

[Eduard]

I'm sorry but I feel like I'm being fooled. If someone who has been programming for many years is still only offering single core engines in 2022, that's poor. It's ok for new developers, but those who program engines for longer should show MP versions.

[Steve Maughan]

But speed with 1 core is stunning 5Mnps in starting position - really!?

5 mnps is fast, but it comes back to the age-old question of how are nodes counted? The conventional way is to increment the node count at the start of the alpha-beta / qsearch procedures. You can increase the nps by counting the number of moves generated. I don't know how Goliath is counting nodes, maybe the author can confirm.

Steve

[Alexander Schmidt]

I'm sorry but I feel like I'm being fooled. If someone who has been programming for many years is still only offering single core engines in 2022, that's poor. It's ok for new developers, but those who program engines for longer should show MP versions.

It's up to the author what kind of things he want to do. Michael is experienced with other neural nets, and didn't know about this new technics in computerchess. After 15 years away from chess he found out about nnue, wanted to try it with LG, and offers it for free. You feel fooled because he didn't first implement multi cores for you? Just leave it, if you don't like it. I like it, and I am really excited to play LG NN!

[Chessqueen]

I'm sorry but I feel like I'm being fooled. If someone who has been programming for many years is still only offering single core engines in 2022, that's poor. It's ok for new developers, but those who program engines for longer should show MP versions.

It's up to the author what kind of things he want to do. Michael is experienced with other neural nets, and didn't know about this new technics in computerchess. After 15 years he found out about nnue, wanted to try it with LG, and offers it for free. You feel fooled because he didn't first implement multi cores for you? Just leave it, if you don't like it. I like it, and I am really excited to play LG NN!

What does Little Goliath 3.15 NN show here for the best move for young 16 years old Pragg against the #2 Indian Player GM Vidit ? Ra8! was the correct move NOT Rook d8?