Draws with Knight odds are possible against a top GM?

Discussion of anything and everything relating to chess playing software and machines.

Moderators: hgm, Harvey Williamson, bob

Forum rules
This textbox is used to restore diagrams posted with the [d] tag before the upgrade.
User avatar
Laskos
Posts: 9323
Joined: Wed Jul 26, 2006 8:21 pm
Full name: Kai Laskos

Draws with Knight odds are possible against a top GM?

Post by Laskos » Tue Feb 19, 2019 12:43 pm

Some three years ago I derived that Stockfish and Komodo at 0.5-1.0 million nodes per move are roughly equal in strength to a top GM (2700+ FIDE Elo) playing at 45 minutes + 15 second increment. Nowadays SF10 probably needs the lower bound, or about 0.5 million nodes per move. And Lc0 t30 needs about 2500 nodes per move to surpass the SF10 at that level. If Lc0 t30 (I used ID32930) is mimicking human play much better than a regular AB CPU engine, then Lc0 t30 at 2500 nodes per move will mimic pretty well a top GM playing a 45' + 15'' game.

Having this "human top GM" (Lc0 t30 at 2500 nodes/move) in hand (the real humans are hard to get), I played SF10 on 4 cores at 15 minutes + 5 seconds time control against Lc0 t30 2500 nodes/move at Knight odds (Stockfish's white Knight b1 removed and then building a small opening book). In 10 games, 8 times Leela (human) prevailed as black, but SF10 got 2 draws as white without a Knight. I guess that using the full time control 45' + 15'', a bit more draws are possible (not by much, though). There is hardly any possibility to Win at Knight odds against a top GM, but some Draws seem possible for a top regular AB engine at 45' + 15'' time control.

Here are the two draws in 10 games SF10 got:


Uri Blass
Posts: 8530
Joined: Wed Mar 08, 2006 11:37 pm
Location: Tel-Aviv Israel

Re: Draws with Knight odds are possible against a top GM?

Post by Uri Blass » Wed Mar 06, 2019 7:52 am

I do not believe that lc0 is mimicking human play much better than a regular AB CPU engine in the first place.
I read that lc0 does not care between different wins and does not care about the fastest win and it is not typical for humans.

User avatar
cdani
Posts: 2104
Joined: Sat Jan 18, 2014 9:24 am
Location: Andorra
Contact:

Re: Draws with Knight odds are possible against a top GM?

Post by cdani » Wed Mar 06, 2019 10:37 pm

Uri Blass wrote:
Wed Mar 06, 2019 7:52 am
I do not believe that lc0 is mimicking human play much better than a regular AB CPU engine in the first place.
I read that lc0 does not care between different wins and does not care about the fastest win and it is not typical for humans.
Anyway is a very interesting experiment, as should be more human like than other alternatives, and lc0 should be overlooking not a few tactics.

Michael Sherwin
Posts: 3041
Joined: Fri May 26, 2006 1:00 am
Location: WY, USA
Full name: Michael Sherwin

Re: Draws with Knight odds are possible against a top GM?

Post by Michael Sherwin » Wed Mar 06, 2019 11:23 pm

Laskos wrote:
Tue Feb 19, 2019 12:43 pm
Some three years ago I derived that Stockfish and Komodo at 0.5-1.0 million nodes per move are roughly equal in strength to a top GM (2700+ FIDE Elo) playing at 45 minutes + 15 second increment. Nowadays SF10 probably needs the lower bound, or about 0.5 million nodes per move. And Lc0 t30 needs about 2500 nodes per move to surpass the SF10 at that level. If Lc0 t30 (I used ID32930) is mimicking human play much better than a regular AB CPU engine, then Lc0 t30 at 2500 nodes per move will mimic pretty well a top GM playing a 45' + 15'' game.

Having this "human top GM" (Lc0 t30 at 2500 nodes/move) in hand (the real humans are hard to get), I played SF10 on 4 cores at 15 minutes + 5 seconds time control against Lc0 t30 2500 nodes/move at Knight odds (Stockfish's white Knight b1 removed and then building a small opening book). In 10 games, 8 times Leela (human) prevailed as black, but SF10 got 2 draws as white without a Knight. I guess that using the full time control 45' + 15'', a bit more draws are possible (not by much, though). There is hardly any possibility to Win at Knight odds against a top GM, but some Draws seem possible for a top regular AB engine at 45' + 15'' time control.

Here are the two draws in 10 games SF10 got:

Are you overestimating SF when giving knight odds? I am no grandmaster. Not even close. I just recently played SF 10 (10 threads) 4 games at QN odds. I was black in all 4 games and my score was 3 draws and one win. The first and last game are here. viewtopic.php?f=11&t=69474&start=30#p785903
I hate if statements. Pawns demand if statements. Therefore I hate pawns.

Uri Blass
Posts: 8530
Joined: Wed Mar 08, 2006 11:37 pm
Location: Tel-Aviv Israel

Re: Draws with Knight odds are possible against a top GM?

Post by Uri Blass » Thu Mar 07, 2019 12:14 am

cdani wrote:
Wed Mar 06, 2019 10:37 pm
Uri Blass wrote:
Wed Mar 06, 2019 7:52 am
I do not believe that lc0 is mimicking human play much better than a regular AB CPU engine in the first place.
I read that lc0 does not care between different wins and does not care about the fastest win and it is not typical for humans.
Anyway is a very interesting experiment, as should be more human like than other alternatives, and lc0 should be overlooking not a few tactics.
I do not think that overlooking tactics is human like style.

I believe that lc0 does not pass the turing test(assuming strong humans look at many games of her) because it does too many tactical errors relative to her chess strength(note that you can use lc0 at small number of nodes per move to avoid not passing the turing test because it is too strong).

Uri Blass
Posts: 8530
Joined: Wed Mar 08, 2006 11:37 pm
Location: Tel-Aviv Israel

Re: Draws with Knight odds are possible against a top GM?

Post by Uri Blass » Thu Mar 07, 2019 12:45 am

I can add that watching leela's losses in the match against stockfish without book I could see tactical mistakes that are clearly not typical human mistakes.

Here is one example



1)For humans it is clear that white has at least a draw by perpetual check so there is no reason to take risks by Kf7 e6+ and it is better to agree to a draw by Kh8 Kh7 when only the queen can give checks so it is a draw.

2)Even a player with fide rating near 2000 like me can calculate Kf7 e6+ Kxe6(or Kg7 Bd4+ with Qh8 mate) Qe7+ Kd5(or Kf5 Qe5+ Kg4 Qe6+ Kf3 Qe2 mate) Qd6+ Ke4 Qe5+ Kf3 Qe2 mate

There are not many lines to calculate.

The blunder of Kf7 is a strong evidence that lc0 is not a human.

lkaufman
Posts: 3647
Joined: Sun Jan 10, 2010 5:15 am
Location: Maryland USA
Contact:

Re: Draws with Knight odds are possible against a top GM?

Post by lkaufman » Thu Mar 07, 2019 5:34 am

Very interesting as always Kai! Some comments:
1. I think Lc0 is better at giving handicaps than Stockfish or Komodo, provided that the network choice and handicap show a win prob. of 3% or more. Network 11248 is great at giving knight odds because it shows a win prob. of around 13%. Ridiculous, but it results in much better play at the handicap than normal engines. Network 41348 gives only around 2.3%, enough to play quite well but perhaps not as well as SF. Some other networks show win prob. of just a bit over 1%, and they play awful, blundering left and right! So it might be interesting to have network 11248 (or any network showing a win prob of say 5% or more) be the entity giving knight odds.
2. Somewhat surprisingly, Lc0 insists, rightly in my opinion, that giving the g1 knight is easier than giving the b1 knight! I know that Morphy and most of the great players of his era usually chose to give the b1 knight (not Lasker though), but I think that this was because the opponents almost always answered 1e4 with 1...e5, rather than 1...d5!, either out of ignorance or perhaps out of some implicit or explicit agreement not to do this. Both 1e4 and 1d4 lead to better positions with optimum play with the g1 knight off than with the b1 knight off. I don't know how you constructed your openings, but the choice of openings seems to have a big effect on results at knight odds.
3. My own experience, playing "slow blitz" games (like 5' + 5" for example) with suitable Lc0 networks (on a 1060 gpu), is that I am lucky to get a draw, at least with the g1 knight handicap. Of course this is a long way from 45' + 15" games, and my current playing strength is a long way below my GM title, but it does lead me to believe that once the problems of bad play at low win probs are worked out, Lc0 on a 2080 might indeed be able to score some points (probably only draws) in a 45' + 15" match with a 2500 FIDE GM. Basically, the difference is that when a normal A/B engine gives knight odds, it is obviously playing for a draw, whereas Lc0 acts as if it expects and intends to win the game right from the start! It doesn't assume that I will see everything that it sees, which makes it far more tricky.
Komodo rules!

User avatar
M ANSARI
Posts: 3399
Joined: Thu Mar 16, 2006 6:10 pm

Re: Draws with Knight odds are possible against a top GM?

Post by M ANSARI » Thu Mar 07, 2019 8:24 am

I think if you wanted to create a network for Lc0 that plays really well at Knight odds is to train it to play without the Knight. I think it will very quickly figure out the best chances in the position and play much stronger than if it was simply playing with a network trained on the normal start position. This is why I also think that Lc0 will soon be unbeatable at Chess 960. If it took Google 4 hours to get a strong network at the start position ... theoretically once hardware catches up you could have a network for each of the Chess 960 positions and if the few weaknesses of Lc0 are sorted out, this would make a formidable Chess 960 engine as it would play each of the start positions equal to how it now plays the classic chess start position.

User avatar
Mike S.
Posts: 1480
Joined: Thu Mar 09, 2006 4:33 am

Re: Draws with Knight odds are possible against a top GM?

Post by Mike S. » Thu Mar 07, 2019 12:37 pm

Uri Blass wrote:
Thu Mar 07, 2019 12:45 am
The blunder of Kf7 is a strong evidence that lc0 is not a human.
I agree. Interesting avoid move position :shock: do you have the complete PGN source of the game? I guess it is from the TCEC bonus event, but I cannot find it. Thanks.
Regards, Mike

Uri Blass
Posts: 8530
Joined: Wed Mar 08, 2006 11:37 pm
Location: Tel-Aviv Israel

Re: Draws with Knight odds are possible against a top GM?

Post by Uri Blass » Thu Mar 07, 2019 4:23 pm


Post Reply