AlphaZero No Castling Chess

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

Moderators: bob, hgm, Harvey Williamson

Forum rules
This textbox is used to restore diagrams posted with the [d] tag before the upgrade.
lkaufman
Posts: 4040
Joined: Sun Jan 10, 2010 5:15 am
Location: Maryland USA
Contact:

Re: AlphaZero No Castling Chess

Post by lkaufman » Sat Dec 21, 2019 8:59 pm

Laskos wrote:
Sat Dec 21, 2019 6:37 pm
lkaufman wrote:
Fri Dec 20, 2019 2:25 am
Laskos wrote:
Thu Dec 19, 2019 2:51 pm
I will try to build now a reasonable book (EPD suite) with Lc0. It doesn't have Contempt, though.


With this trapezoidal noise for Lc0, limiting the bad moves to no more than 2% offset from the best move (win probability)

{
"name" : "Temperature",
"value" : "0.60"
},
{
"name" : "TempDecayMoves",
"value" : "9"
},
{
"name" : "TempValueCutoff",
"value" : "2.00"
},
{
"name" : "TempCutoffMove",
"value" : 6
}

I am getting a small opening suite of 6-movers built with T40 Lc0, 266 positions

http://s000.tinyupload.com/?file_id=900 ... 8031526635

I will try to build a larger book by widening a bit the trapezoid at its narrow end-side, and by allowing an offset of 3% win probability instead of 2%.
I have built a larger 6-mover opening suite with Lc0 with 476 positions:
https://file.io/w77hc1

But I think the openings in it are discrediting Black's chances. Lc0 has an "in innate" Contempt for the side to move (I have shown that in the past), which in openings means that White plays well, while Black is wrongly aggressive, instead of going to resistance for a draw. Look at the openings, aren't many of them exhibiting that? As expected the White performance goes up with wrong for Black Lc0 openings:

1000 games at 30 + 0.3s
SF_9 vs K_131 (White Contempt = 50)
52.5%

1000 games
K_AB (20+0.2s) vs K_MCTS (30+0.3s) both White Contempt = 50
55.1%

1000 games, 1000 nodes/move
Lc0 T40 vs Lc0 T30
52.1%

==================

53.2% White performance (Armageddon scoring) in 3000 games with openings built using Lc0.

I think the only reliable way to build representative openings for this unbalanced variant is by using Komodo with its Colored Contempt. Contempt usage IS important from the start of the game.
I tried your new (smaller) book at a long time control, 10' + 6", and after 102 games got 54.9% White wins. Small sample, but it's good to know that the White plus is probably due to the book rather than the TC. The openings looked very "normal", as in ones that we might see in normal chess, which probably reflects the NN bias for moves that are "normally" best. Black should instead seek defenses suitable for the queenside only castling situation, which are not a random sample of normal openings. The other problem I saw is that the same defenses kept coming up again and again, with just a tiny difference each time. For example the Stonewall Dutch was played many times. So in short I agree with you.
Komodo rules!

User avatar
Laskos
Posts: 10175
Joined: Wed Jul 26, 2006 8:21 pm
Full name: Kai Laskos

Re: AlphaZero No Castling Chess

Post by Laskos » Sat Dec 21, 2019 9:43 pm

lkaufman wrote:
Sat Dec 21, 2019 8:59 pm
Laskos wrote:
Sat Dec 21, 2019 6:37 pm
lkaufman wrote:
Fri Dec 20, 2019 2:25 am
Laskos wrote:
Thu Dec 19, 2019 2:51 pm
I will try to build now a reasonable book (EPD suite) with Lc0. It doesn't have Contempt, though.


With this trapezoidal noise for Lc0, limiting the bad moves to no more than 2% offset from the best move (win probability)

{
"name" : "Temperature",
"value" : "0.60"
},
{
"name" : "TempDecayMoves",
"value" : "9"
},
{
"name" : "TempValueCutoff",
"value" : "2.00"
},
{
"name" : "TempCutoffMove",
"value" : 6
}

I am getting a small opening suite of 6-movers built with T40 Lc0, 266 positions

http://s000.tinyupload.com/?file_id=900 ... 8031526635

I will try to build a larger book by widening a bit the trapezoid at its narrow end-side, and by allowing an offset of 3% win probability instead of 2%.
I have built a larger 6-mover opening suite with Lc0 with 476 positions:
https://file.io/w77hc1

But I think the openings in it are discrediting Black's chances. Lc0 has an "in innate" Contempt for the side to move (I have shown that in the past), which in openings means that White plays well, while Black is wrongly aggressive, instead of going to resistance for a draw. Look at the openings, aren't many of them exhibiting that? As expected the White performance goes up with wrong for Black Lc0 openings:

1000 games at 30 + 0.3s
SF_9 vs K_131 (White Contempt = 50)
52.5%

1000 games
K_AB (20+0.2s) vs K_MCTS (30+0.3s) both White Contempt = 50
55.1%

1000 games, 1000 nodes/move
Lc0 T40 vs Lc0 T30
52.1%

==================

53.2% White performance (Armageddon scoring) in 3000 games with openings built using Lc0.

I think the only reliable way to build representative openings for this unbalanced variant is by using Komodo with its Colored Contempt. Contempt usage IS important from the start of the game.
I tried your new (smaller) book at a long time control, 10' + 6", and after 102 games got 54.9% White wins. Small sample, but it's good to know that the White plus is probably due to the book rather than the TC. The openings looked very "normal", as in ones that we might see in normal chess, which probably reflects the NN bias for moves that are "normally" best. Black should instead seek defenses suitable for the queenside only castling situation, which are not a random sample of normal openings. The other problem I saw is that the same defenses kept coming up again and again, with just a tiny difference each time. For example the Stonewall Dutch was played many times. So in short I agree with you.
Here is the file of what Komodo White Contempt 50 or 75 considers as the more often occurring 5-mover openings (3 or more times occurring). The number of occurrences in games is on the right side of the FEN string in the EPD file (giving the order of popularity).
https://file.io/m2YIEW

Be careful, for example LittleBlitzer will refuse to read these FEN strings. Now testing in 1000 games from this file of more robust but fewer openings SF_9 (Contempt = 0) against Komodo (White Contempt = 50). After about 300 games, about 49% White scoring. I will report tomorrow the final result.

Paloma
Posts: 915
Joined: Thu Dec 25, 2008 8:07 pm

Re: AlphaZero No Castling Chess

Post by Paloma » Sun Dec 22, 2019 1:14 am

"404 Page not found"

User avatar
Laskos
Posts: 10175
Joined: Wed Jul 26, 2006 8:21 pm
Full name: Kai Laskos

Re: AlphaZero No Castling Chess

Post by Laskos » Sun Dec 22, 2019 7:22 am

Paloma wrote:
Sun Dec 22, 2019 1:14 am
"404 Page not found"
I wii re-post it today when I will be at my PC. These are fast expiring links.

User avatar
Laskos
Posts: 10175
Joined: Wed Jul 26, 2006 8:21 pm
Full name: Kai Laskos

Re: AlphaZero No Castling Chess

Post by Laskos » Sun Dec 22, 2019 6:55 pm

Laskos wrote:
Sun Dec 22, 2019 7:22 am
Paloma wrote:
Sun Dec 22, 2019 1:14 am
"404 Page not found"
I wii re-post it today when I will be at my PC. These are fast expiring links.

Here is the file with the most popular 5-movers (appearing at least 3 times in the games) according to Komodo. The number of occurrences (popularity) is written after each FEN string:
https://gofile.io/?c=0bCAWV

I got a bit surprising result SF_9 (Contempt = 0) against Komodo (White Contempt = 50) at 30s + 0.3s in 1000 games from this smaller but hopefully better quality suite:

White scoring:

+458 =455 -87 68.5%
White score (Armageddon): 45.8%

Will test now T40 against T30 with Lc0.

User avatar
Laskos
Posts: 10175
Joined: Wed Jul 26, 2006 8:21 pm
Full name: Kai Laskos

Re: AlphaZero No Castling Chess

Post by Laskos » Sun Dec 22, 2019 9:39 pm

Laskos wrote:
Sun Dec 22, 2019 6:55 pm
Laskos wrote:
Sun Dec 22, 2019 7:22 am
Paloma wrote:
Sun Dec 22, 2019 1:14 am
"404 Page not found"
I wii re-post it today when I will be at my PC. These are fast expiring links.

Here is the file with the most popular 5-movers (appearing at least 3 times in the games) according to Komodo. The number of occurrences (popularity) is written after each FEN string:
https://gofile.io/?c=0bCAWV

I got a bit surprising result SF_9 (Contempt = 0) against Komodo (White Contempt = 50) at 30s + 0.3s in 1000 games from this smaller but hopefully better quality suite:

White scoring:

+458 =455 -87 68.5%
White score (Armageddon): 45.8%

Will test now T40 against T30 with Lc0.
The same 162-positions suite.

1000 games
1000 nodes/move t40 vs t30

White scoring:

+522 =383 -95 71.3%
White score (Armageddon): 52.2%

Large variation from AB engines score. Maybe the small sample of only 162 openings is introducing a large systematic error? I wouldn't say so, as some variety is introduced by engines themselves, not only openings. Will check now with K_AB vs K_MCTS.

User avatar
Laskos
Posts: 10175
Joined: Wed Jul 26, 2006 8:21 pm
Full name: Kai Laskos

Re: AlphaZero No Castling Chess

Post by Laskos » Mon Dec 23, 2019 8:48 am

And the last test from this short 162 position suite, 1000 games

Komodo_AB at 20 + 0.2s vs Komodo_MCTS at 30 + 0.3s (both White Contempt = 50):

White scoring:
+500 =441 -59 72.0%
White score (Armageddon): 50.0%

So, in 3,000 games with different combinations of engines the average score is 49.3%.

Post Reply