Chess variants made with help from alpha zero article

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.
Raphexon
Posts: 299
Joined: Sun Mar 17, 2019 11:00 am
Full name: Henk Drost

Re: Chess variants made with help from alpha zero article

Post by Raphexon » Mon Sep 14, 2020 2:21 pm

Laskos wrote:
Mon Sep 14, 2020 1:49 pm
Raphexon wrote:
Mon Sep 14, 2020 12:23 pm
The start position of Capablanca is relatively garbage.
Testing it with Fairy-Stockfish gave me a fairly hefty opening advantage at 60+0.6.
White vs Black: +952 -438 =610 [0.628] 2000
There are also unprotected pawns in the beginning.


I've also tested Gothic Chess. (Capablanca but different start position)
Which was a lot more balanced: White vs black = +360 -291 =338 [0.529]
More balanced than standard chess too.
What openings did you use?
From start with 4 threads...
So there might be artifacts... But atleast SMP prevents it from being totally deterministic. It does have a bias towards certain opening moves but games did quickly diverge from having eyeballed it.
When I'm home I can share PGNs.

Since there's no real opening theory for it I'd have to create an opening book myself.
Which would still add artifacts because I'd be using SF MultiPV to pick viable moves.

User avatar
Laskos
Posts: 10620
Joined: Wed Jul 26, 2006 8:21 pm
Full name: Kai Laskos

Re: Chess variants made with help from alpha zero article

Post by Laskos » Mon Sep 14, 2020 2:55 pm

Raphexon wrote:
Mon Sep 14, 2020 2:21 pm
Laskos wrote:
Mon Sep 14, 2020 1:49 pm
Raphexon wrote:
Mon Sep 14, 2020 12:23 pm
The start position of Capablanca is relatively garbage.
Testing it with Fairy-Stockfish gave me a fairly hefty opening advantage at 60+0.6.
White vs Black: +952 -438 =610 [0.628] 2000
There are also unprotected pawns in the beginning.


I've also tested Gothic Chess. (Capablanca but different start position)
Which was a lot more balanced: White vs black = +360 -291 =338 [0.529]
More balanced than standard chess too.
What openings did you use?
From start with 4 threads...
So there might be artifacts... But atleast SMP prevents it from being totally deterministic. It does have a bias towards certain opening moves but games did quickly diverge from having eyeballed it.
When I'm home I can share PGNs.

Since there's no real opening theory for it I'd have to create an opening book myself.
Which would still add artifacts because I'd be using SF MultiPV to pick viable moves.
Have 4 - 8 reasonable 2ply openings, that will improve the variability of the openings and decrease the artifacts by an order of magnitude. Then play with 2 or 4 threads.

Raphexon
Posts: 299
Joined: Sun Mar 17, 2019 11:00 am
Full name: Henk Drost

Re: Chess variants made with help from alpha zero article

Post by Raphexon » Mon Sep 14, 2020 4:49 pm

Laskos wrote:
Mon Sep 14, 2020 2:55 pm
Raphexon wrote:
Mon Sep 14, 2020 2:21 pm
Laskos wrote:
Mon Sep 14, 2020 1:49 pm
Raphexon wrote:
Mon Sep 14, 2020 12:23 pm
The start position of Capablanca is relatively garbage.
Testing it with Fairy-Stockfish gave me a fairly hefty opening advantage at 60+0.6.
White vs Black: +952 -438 =610 [0.628] 2000
There are also unprotected pawns in the beginning.


I've also tested Gothic Chess. (Capablanca but different start position)
Which was a lot more balanced: White vs black = +360 -291 =338 [0.529]
More balanced than standard chess too.
What openings did you use?
From start with 4 threads...
So there might be artifacts... But atleast SMP prevents it from being totally deterministic. It does have a bias towards certain opening moves but games did quickly diverge from having eyeballed it.
When I'm home I can share PGNs.

Since there's no real opening theory for it I'd have to create an opening book myself.
Which would still add artifacts because I'd be using SF MultiPV to pick viable moves.
Have 4 - 8 reasonable 2ply openings, that will improve the variability of the openings and decrease the artifacts by an order of magnitude. Then play with 2 or 4 threads.
I'm going to follow your advice and create a small opening book.
There are 28 opening moves for Capablanca with plenty of viable ones:
Going to run till depth 25, select a few and run those until depth 25 too.

Think I'm going to pick all moves within 50cp from best.

Code: Select all

<fairy-stockfish-largeboard_x86-64-bmi2(0): info depth 21 seldepth 35 multipv 1 score cp 91 nodes 150607443 nps 231621 hashfull 937 tbhits 0 time 650232 pv i1h3 i8h6 e2e4 f7f5 b1c3 e7e5 f2f4 e5f4 c1d3 d7d6 d3f4 g8f7 g1f2 b8c6 d2d3 c6e5 h3g5 d8g5 f4g5 f5e4
<fairy-stockfish-largeboard_x86-64-bmi2(0): info depth 21 seldepth 35 multipv 2 score cp 73 nodes 150607443 nps 231621 hashfull 937 tbhits 0 time 650232 pv e2e4 d7d6 i1h3 i8h6 f2f4 f7f5 b1c3 e7e5 d2d3 b8c6 h1f2 h8f7 e4f5 c8f5 c3e4 f5g6 c2c3 e5f4 c1f4 g6f4 h3f4 e8e5 g2g3 d6d5 d1b3
<fairy-stockfish-largeboard_x86-64-bmi2(0): info depth 21 seldepth 35 multipv 3 score cp 71 nodes 150607443 nps 231621 hashfull 937 tbhits 0 time 650232 pv f2f4 f7f5 d2d4 i8h6 c1d3 d7d5 b1d2 c8d6 d2f3 b8d7 i1h3 d7f6 f3e5 f6e4 e2e3 e7e6 c2c4 c7c5 c4d5 e6d5 d4c5 e4c5
<fairy-stockfish-largeboard_x86-64-bmi2(0): info depth 21 seldepth 36 multipv 4 score cp 70 nodes 150607443 nps 231621 hashfull 937 tbhits 0 time 650232 pv d2d4 i8h6 i1h3 f7f5 f2f4 d7d5 b1d2 b8d7 d2f3 e7e6 c1d3 c8d6 f3e5 d8e7 e2e3 c7c5 e5d7 e8d7 d4c5 d6e4 d3e5 d7b5 d1e2
<fairy-stockfish-largeboard_x86-64-bmi2(0): info depth 21 seldepth 38 multipv 5 score cp 58 nodes 150607443 nps 231621 hashfull 937 tbhits 0 time 650232 pv d2d3 i8h6 e2e4 d7d6 i1h3 f7f5 b1c3 b8c6 f2f4 e7e5 e4f5 c8f5 c3e4 h6g4 d1g4 f5g4 h1f2 g4f2 e1f2 h8f7 h3g5 d8g5 e4g5 f7f6 g5e4 f6f4 c1f4 e5f4 f2f4 e8f7 f4f7 g8f7 e4g5
<fairy-stockfish-largeboard_x86-64-bmi2(0): info depth 21 seldepth 37 multipv 6 score cp 32 nodes 150607443 nps 231621 hashfull 937 tbhits 0 time 650232 pv j2j4 i8h6 f2f4 f7f5 i1h3 c8d6 d2d4 h8g6 h1g3 g8d5 b1c3 d5c4 g3g6 h7g6 c3b1 h6g4 h3i5 g4i5 j4i5 e7e6
<fairy-stockfish-largeboard_x86-64-bmi2(0): info depth 21 seldepth 39 multipv 7 score cp 30 nodes 150607443 nps 231621 hashfull 937 tbhits 0 time 650232 pv b1c3 e7e5 d2d3 i8h6 f2f4 d7d6 i1h3 b8c6 e2e4 f7f5 e1f2 c8b6 f2i5 b6d7 d1h5 g8f7 h5f7 h8f7 i5i7 j8i8 i7h7 g7g5 h3g5 d8g5 h7f7 h6f7 f4g5 f5e4 c3e4 i8i2
<fairy-stockfish-largeboard_x86-64-bmi2(0): info depth 21 seldepth 32 multipv 8 score cp 24 nodes 150607443 nps 231621 hashfull 937 tbhits 0 time 650232 pv c2c4 d7d6 i1h3 e7e5 d2d3 i8h6 f2f4 f7f5 f4e5 d6e5 e2e4 f5f4 c1b3 b8c6 g1c5 d8e7 b1c3 c8e6 c5e7 e8e7 c3d5 e7d6 c4c5
<fairy-stockfish-largeboard_x86-64-bmi2(0): info depth 21 seldepth 39 multipv 9 score cp 12 nodes 150607443 nps 231621 hashfull 937 tbhits 0 time 650232 pv e2e3 i8h6 f2f4 e7e5 f4e5 e8e5 b1c3 f7f5 c1e2 c8d6 i1h3 g7g5 e2g3 e5e6 d2d3 b8c6 g1f2 h6g4 g3d6 c7d6 d1f3 h8g6
<fairy-stockfish-largeboard_x86-64-bmi2(0): info depth 21 seldepth 31 multipv 10 score cp 6 nodes 150607443 nps 231621 hashfull 937 tbhits 0 time 650232 pv g2g4 d7d6 h2h3 f7f5 g4f5 c8f5 i1g2 e7e5 d2d3 f5h6 b1c3 h6c1 a1c1 i8h6 e2e3 b8d7 h1g3 h8g6 f2f4 j7j6 e3e4 c7c6
<fairy-stockfish-largeboard_x86-64-bmi2(0): info depth 21 seldepth 37 multipv 11 score cp 3 nodes 150607443 nps 231621 hashfull 937 tbhits 0 time 650232 pv a2a4 d7d5 i1h3 i8h6 d2d4 f7f5 f2f4 b8d7 e2e3 c8d6 b1d2 e7e6 c1e2 h6g4 h3g5 g4f6 g5f3 c7c5 f3e5 c5c4 h2h3
<fairy-stockfish-largeboard_x86-64-bmi2(0): info depth 21 seldepth 31 multipv 12 score cp 0 nodes 150607443 nps 231621 hashfull 937 tbhits 0 time 650232 pv j2j3 d7d6 i1h3 e7e5 e2e4 f7f5 f2f4 b8c6 d2d3 i8h6 b1c3 h8f7 c1b3 f7d7 b3c1 d7f7
<fairy-stockfish-largeboard_x86-64-bmi2(0): info depth 21 seldepth 33 multipv 13 score cp -9 nodes 150607443 nps 231621 hashfull 937 tbhits 0 time 650232 pv f2f3 d7d6 i1h3 f7f5 d2d4 i8h6 c2c3 e7e5 d1b3 g8b3 c1b3 c8e6 d4e5 e6b3 a2b3 d6e5 a1a7 a8a7 g1a7 b8d7 h1f2 h8i6 i2i3 d8f6 e2e4 f8i8
<fairy-stockfish-largeboard_x86-64-bmi2(0): info depth 21 seldepth 35 multipv 14 score cp -19 nodes 150607443 nps 231621 hashfull 937 tbhits 0 time 650232 pv h2h3 f7f5 g2g4 f5g4 d2d3 i8h6 h3g4 d7d6 i1h3 h6g4 f2f3 g4h6 e2e4 g8f7 f3f4 e7e5 g1h2 f7i4 f4f5 g7g6 e1i5 g6f5 i5i7
<fairy-stockfish-largeboard_x86-64-bmi2(0): info depth 21 seldepth 38 multipv 15 score cp -33 nodes 150607443 nps 231621 hashfull 937 tbhits 0 time 650232 pv h1g3 e7e5 d2d4 e5d4 c1i7 i8h6 i7h8 j8h8 c2c3 c8d6 c3d4 f7f5 b1c3 g8c4 d1b3 c4b3 a2b3 h8i8 d4d5 h6g4 i1h3
<fairy-stockfish-largeboard_x86-64-bmi2(0): info depth 20 seldepth 28 multipv 16 score cp 4 nodes 150607443 nps 231621 hashfull 937 tbhits 0 time 650232 pv b1a3 i8h6 i1h3 e7e5 e2e4 d7d6 d2d3 f7f5 f2f4 b8c6 h1f2 h8f7 h3g5 f7f6 f4e5 e8e5 d3d4 e5e7 d4d5 c6e5 c1f4 f5e4 g5e4
<fairy-stockfish-largeboard_x86-64-bmi2(0): info depth 20 seldepth 28 multipv 17 score cp 0 nodes 150607443 nps 231621 hashfull 937 tbhits 0 time 650232 pv c1d3 d7d6
<fairy-stockfish-largeboard_x86-64-bmi2(0): info depth 20 seldepth 38 multipv 18 score cp -2 nodes 150607443 nps 231621 hashfull 937 tbhits 0 time 650232 pv a2a3 e7e5
<fairy-stockfish-largeboard_x86-64-bmi2(0): info depth 20 seldepth 33 multipv 19 score cp -9 nodes 150607443 nps 231621 hashfull 937 tbhits 0 time 650232 pv c2c3 e7e5 i1h3 i8h6 e2e4 f7f5 f2f4 f5e4 d2d3 d7d5 h1f2 h8f7 d3e4 c8d6 c1d3 e5f4 e4e5 e8e5 d3e5 d6e5 f2f3 e5d6 b1d2 d8e7
<fairy-stockfish-largeboard_x86-64-bmi2(0): info depth 20 seldepth 30 multipv 20 score cp -10 nodes 150607443 nps 231621 hashfull 937 tbhits 0 time 650232 pv i2i4 e7e5 i1h3 h7h6 e2e4 d7d6 d2d3 f7f5 f2f4 f5e4 f4e5 e8f7 g1f2 b8c6 h1g3 c6e5 g3e4 c8f5 f1i1 c7c5 b1c3 f5e4 e1e4 g8h7
<fairy-stockfish-largeboard_x86-64-bmi2(0): info depth 20 seldepth 30 multipv 21 score cp -14 nodes 150607443 nps 231621 hashfull 937 tbhits 0 time 650232 pv b2b3 e7e5 e2e4 d7d6 i1h3 f7f5 f2f4 e5f4 c1d3 i8h6 d3f4 f5e4 b1c3 d8f6 d1h5 e8c6 h3g5 c8i2 e1e4 c6e4 g5e4 i2h1 j1h1
<fairy-stockfish-largeboard_x86-64-bmi2(0): info depth 20 seldepth 33 multipv 22 score cp -21 nodes 150607443 nps 231621 hashfull 937 tbhits 0 time 650232 pv h2h4 i8h6 f2f4 d7d5 i1h3 f7f5 d2d4 b8d7 b1d2 c7c5 d4c5 e7e5 f4e5 d7e5 d2f3 e5f3 g2f3 h8i6 i2i3 g8f7 c1f4 f8i8 e2e3 f7i4 d1e2 d8f6 e1b4
<fairy-stockfish-largeboard_x86-64-bmi2(0): info depth 20 seldepth 29 multipv 23 score cp -26 nodes 150607443 nps 231621 hashfull 937 tbhits 0 time 650232 pv b2b4 i8h6 i1h3 e7e5 d2d3 f7f5 c2c4 d8f6 b1c3 d7d6 f2f4 b8d7 e2e4 e5f4 h3f4 f5e4 h1f2 d7e5 c3e4
<fairy-stockfish-largeboard_x86-64-bmi2(0): info depth 20 seldepth 32 multipv 24 score cp -33 nodes 150607443 nps 231621 hashfull 937 tbhits 0 time 650232 pv g2g3 e7e5 d2d3 i8h6 i1h3 f7f5 f2f4 e5e4 c2c4 d7d6 b1c3 h6g4 h3f2 e4d3 e2d3 e8e1 f1e1 g4h6 c1b3 b8c6 f2h3
<fairy-stockfish-largeboard_x86-64-bmi2(0): info depth 20 seldepth 37 multipv 25 score cp -50 nodes 150607443 nps 231621 hashfull 937 tbhits 0 time 650232 pv i2i3 f7f5 d2d3 i8h6 c2c4 e7e5 i1h3 g8e6 f2f4 d7d6 b1c3 c7c6 f4e5 d6e5 g1c5 d8e7 c5e7 c8e7 e2e4 h8g6 h1f2
<fairy-stockfish-largeboard_x86-64-bmi2(0): info depth 20 seldepth 31 multipv 26 score cp -59 nodes 150607443 nps 231621 hashfull 937 tbhits 0 time 650232 pv i1j3 e7e5 h2h3 i8h6 e2e4 d7d6 b1c3 f7f5 d2d3 b8c6 e4f5 c8f5 f2f4 f5e6 h1f2 h8f7 g1h2 e6d4 f1i1 d4f2 e1f2 f7f6
<fairy-stockfish-largeboard_x86-64-bmi2(0): info depth 20 seldepth 28 multipv 27 score cp -104 nodes 150607443 nps 231621 hashfull 937 tbhits 0 time 650232 pv c1b3 d7d6 i1h3 e7e5 c2c4 g7g5 f2f3 f7f5 b1c3 i8h6 d2d4 g5g4 h3j4 h8g6 d4d5 g8f7 j4h5 f8i8 e2e3 c7c6 e3e4 f5f4 f3g4 c6d5 c4d5
<fairy-stockfish-largeboard_x86-64-bmi2(0): info depth 20 seldepth 30 multipv 28 score cp -111 nodes 150607443 nps 231621 hashfull 937 tbhits 0 time 650232 pv h1i3 e7e5 i3h1 i8h6 d2d3 c8d6 e2e4 f7f5 b1c3 b8c6 h2h3 f5f4 g2g3 g7g5 g3f4 e5f4 f2f3 c6e5 i1g2 c7c6 h1i3

jp
Posts: 1392
Joined: Mon Apr 23, 2018 5:54 am

Re: Chess variants made with help from alpha zero article

Post by jp » Tue Sep 15, 2020 1:56 am

Raphexon wrote:
Mon Sep 14, 2020 12:23 pm
The start position of Capablanca is relatively garbage.
Testing it with Fairy-Stockfish gave me a fairly hefty opening advantage at 60+0.6.
White vs Black: +952 -438 =610 [0.628] 2000
There are also unprotected pawns in the beginning.

I've also tested Gothic Chess. (Capablanca but different start position)
Which was a lot more balanced: White vs black = +360 -291 =338 [0.529]
More balanced than standard chess too.
Do you know the opening advantage for Shogi, Xiangqi, checkers, backgammon, etc. (in terms of win percentage of each side)?

Raphexon
Posts: 299
Joined: Sun Mar 17, 2019 11:00 am
Full name: Henk Drost

Re: Chess variants made with help from alpha zero article

Post by Raphexon » Tue Sep 15, 2020 7:34 am

jp wrote:
Tue Sep 15, 2020 1:56 am
Raphexon wrote:
Mon Sep 14, 2020 12:23 pm
The start position of Capablanca is relatively garbage.
Testing it with Fairy-Stockfish gave me a fairly hefty opening advantage at 60+0.6.
White vs Black: +952 -438 =610 [0.628] 2000
There are also unprotected pawns in the beginning.

I've also tested Gothic Chess. (Capablanca but different start position)
Which was a lot more balanced: White vs black = +360 -291 =338 [0.529]
More balanced than standard chess too.
Do you know the opening advantage for Shogi, Xiangqi, checkers, backgammon, etc. (in terms of win percentage of each side)?
Shogi's opening advantage is minimal.
The Japanese Shogi federation even had a year where second player (white) won more than the first player.

https://www.chess.com/news/view/white-p ... vents-7702

No idea about the rest.

GregNeto
Posts: 34
Joined: Thu Sep 08, 2016 6:07 am

Re: Chess variants made with help from alpha zero article

Post by GregNeto » Tue Sep 15, 2020 9:06 am

A nice variant is kq-chess: swap the white king and queen positions (king on d1, queen on e1); white short castle goes to the queenside (king on b1, rook on c1), white long castle to the kingside (king f1, rook e1).
A lot of opposite castling should lead to more attacking games and fewer draws. Totally new openings, but still the chess that we all know and love. And its not just one of the 960 positions!
I played this variant a lot with friends, we enjoy it.
Should be easy to implement and test for the expert programmers.

Have fun
Greg

jp
Posts: 1392
Joined: Mon Apr 23, 2018 5:54 am

Re: Chess variants made with help from alpha zero article

Post by jp » Wed Sep 16, 2020 3:00 am

Raphexon wrote:
Tue Sep 15, 2020 7:34 am
jp wrote:
Tue Sep 15, 2020 1:56 am
Do you know the opening advantage for Shogi, Xiangqi, checkers, backgammon, etc. (in terms of win percentage of each side)?
Shogi's opening advantage is minimal.
The Japanese Shogi federation even had a year where second player (white) won more than the first player.

https://www.chess.com/news/view/white-p ... vents-7702

No idea about the rest.
Apparently, an unnamed online Xiangqi database of 70,000 expert games shows:

Red win 37.78%
Black win 27.90%
Draw 34.32%.

Vinvin
Posts: 4752
Joined: Thu Mar 09, 2006 8:40 am
Full name: Vincent Lejeune

Re: Chess variants made with help from alpha zero article

Post by Vinvin » Wed Sep 16, 2020 9:18 am

jp wrote:
Wed Sep 16, 2020 3:00 am
Raphexon wrote:
Tue Sep 15, 2020 7:34 am
jp wrote:
Tue Sep 15, 2020 1:56 am
Do you know the opening advantage for Shogi, Xiangqi, checkers, backgammon, etc. (in terms of win percentage of each side)?
Shogi's opening advantage is minimal.
The Japanese Shogi federation even had a year where second player (white) won more than the first player.

https://www.chess.com/news/view/white-p ... vents-7702

No idea about the rest.
Apparently, an unnamed online Xiangqi database of 70,000 expert games shows:

Red win 37.78%
Black win 27.90%
Draw 34.32%.
Are there some statistics from the computers ?
http://wdoor.c.u-tokyo.ac.jp/shogi/logs ... dgate.html
https://en.wikipedia.org/wiki/Computer_shogi#Floodgate

Computers improved by more than 1000 Elo since they become stronger than top humans (in 2014).

Post Reply