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.
Alayan
Posts: 394
Joined: Tue Nov 19, 2019 7:48 pm
Full name: Alayan Feh

Re: Chess variants made with help from alpha zero article

Post by Alayan » Thu Sep 10, 2020 11:23 pm

mmt wrote:
Thu Sep 10, 2020 6:59 am
Alayan wrote:
Thu Sep 10, 2020 6:42 am
13,94% decisive games vs 11,80% decisive games at 1s/move ; and 2,9% vs 2,1% at 1m/move, that's a very substantial increase. Chess is a drawish game and for a given strength NN self-play is extremely drawish.

In a tournament like TCEC, 25% decisive games instead of 20% would be a massive difference.
I don't think it's enough of an increase to make it a variant that people or programmers would switch to unless there are other reasons.
The point is that it plays very similarly to classical chess. Opening theory doesn't change much, there are some significant change to endgame theory but it's not overwhelming.

So it would be the most likely to be adopted in human chess, and being adopted by human players is the single most important factor to get more than a handful of engine authors to switch.
mmt wrote:
Thu Sep 10, 2020 8:18 am
Raphexon wrote:
Thu Sep 10, 2020 7:36 am
Yes, 38% more wins at 1m/move is a nice start.
Or looking at this from another perspective, you've reduced the number of draws by less than 1%.
98% draws is not what's the situation is like either in human chess or in competitive engine chess with forced opening lines. A risk-averse NN playing itself is worst case scenario.

For the coming years, opening selection will still be able to delay draw death. Last SuFi had 17/50 decisive game pairs. Stalemate=win would delay it more.

But to dramatically decrease draws among strong engines, you need to toss out classical chess and go with some big board variant like Raphexeon suggested. You won't get human players to switch to it anyway so why keep classical chess flaws.

Armageddon variants theoretically work, but are much less appealing. Removing the drawing margin theoretically increase skill ceiling, but that's compared to start position chess. Compared to "TCEC chess" with forced openings that try to be close to the draw/win frontier, you'll be able to display 2x bigger elo spread because you mark draws by the weak side as wins, but really complexity is equivalent. Take NBSC as start position, score draws just as draws, how is it in any way an inferior game to NBSC Armageddon, assuming reverse games are played which is always true in engine chess ? Stefan even went the whole way to score black wins as +2, just keep classical scoring at this point, it will reward those black wins automatically !

I don't see any reason to believe that Armageddon chess variants will be able to delay the "stronger side wins all games" issue longer than "TCEC chess" will.

the_real_greco
Posts: 45
Joined: Thu Sep 19, 2019 3:55 am
Full name: Andy!

Re: Chess variants made with help from alpha zero article

Post by the_real_greco » Fri Sep 11, 2020 1:03 am

Man. "AI ruined chess" seems harsh. I wish DeepMind would just leave chess alone and go bother someone else.

jmartus
Posts: 207
Joined: Sun May 16, 2010 12:50 am

Re: Chess variants made with help from alpha zero article

Post by jmartus » Fri Sep 11, 2020 1:28 am

Did you even read the article?

Collingwood
Posts: 18
Joined: Sat Nov 09, 2019 2:24 pm
Full name: Colin Wood

Re: Chess variants made with help from alpha zero article

Post by Collingwood » Fri Sep 11, 2020 3:13 am

the_real_greco wrote:
Fri Sep 11, 2020 1:03 am
I wish DeepMind would just leave chess alone and go bother someone else.
DeepMind's problem is that they don't do anything really original or fundamental, but they don't do anything practical enough to make money for Google either. They are stuck doing stuff they can hype in the popular media, and chess has always been a good vehicle for corporate propaganda.

mmt
Posts: 274
Joined: Sun Aug 25, 2019 6:33 am
Full name: .

Re: Chess variants made with help from alpha zero article

Post by mmt » Fri Sep 11, 2020 3:14 am

Alayan wrote:
Thu Sep 10, 2020 11:23 pm
The point is that it plays very similarly to classical chess. Opening theory doesn't change much, there are some significant change to endgame theory but it's not overwhelming.

So it would be the most likely to be adopted in human chess, and being adopted by human players is the single most important factor to get more than a handful of engine authors to switch.
We don't know what effect the no-stalemate variant would have on human GM draws, though. It could be minimal. Weak players have few draws in their games so that doesn't matter much. Maybe it's possible to go through the database of highly rated human games to look at endings and try to estimate the impact. We do know that it won't be anywhere near to what armageddon scoring with no-black-castling can do and that's still not a radical change except for memorized openings.
Alayan wrote:
Thu Sep 10, 2020 11:23 pm
98% draws is not what's the situation is like either in human chess or in competitive engine chess with forced opening lines. A risk-averse NN playing itself is worst case scenario.
I don't see a reason to think that Alpha Zero is risk-averse. The percentage of draws is 53% for players with Elo above 2600 in classical games (https://en.chessbase.com/post/has-the-n ... -increased). Maybe I'm not up-to-date but I haven't seen any forced-opening-lines human chess tournaments and as you said, we mostly need humans to adopt it. I don't think they would - you're basically taking away a part of the game.
Alayan wrote:
Thu Sep 10, 2020 11:23 pm
But to dramatically decrease draws among strong engines, you need to toss out classical chess and go with some big board variant like Raphexeon suggested. You won't get human players to switch to it anyway so why keep classical chess flaws.
Right, I don't see chess players switching to something dramatically different like this. I remember when in table tennis they increased the ball size. It required an adjustment in the playing style but it was still table tennis. A change can work and if you're making one, you might as well eliminate draws completely instead of delaying the draw death.
Alayan wrote:
Thu Sep 10, 2020 11:23 pm
Armageddon variants theoretically work, but are much less appealing. Removing the drawing margin theoretically increase skill ceiling, but that's compared to start position chess. Compared to "TCEC chess" with forced openings that try to be close to the draw/win frontier, you'll be able to display 2x bigger elo spread because you mark draws by the weak side as wins, but really complexity is equivalent.
As I wrote earlier, I think forced-openings are a non-starter for humans, so we should compare to classical chess. Making computer games more interesting to watch by making every game a win instead of just half of them in the forced-opening variant is a huge difference.
Alayan wrote:
Thu Sep 10, 2020 11:23 pm
Take NBSC as start position, score draws just as draws, how is it in any way an inferior game to NBSC Armageddon, assuming reverse games are played which is always true in engine chess ?
1. Reverse games would require twice as much time as regular games for humans
2. People don't like draws. It's human psychology to prefer a winner/loser scenario and this theme has been true in sports with overtimes.
3. It allows black to "win" instead of just fighting to draw (assuming very good players). Makes it appear more fair and allows one-time games.
Alayan wrote:
Thu Sep 10, 2020 11:23 pm
I don't see any reason to believe that Armageddon chess variants will be able to delay the "stronger side wins all games" issue longer than "TCEC chess" will.
As I mentioned, the comparison should be to classical chess, not to chess without having to play openings.

Alayan
Posts: 394
Joined: Tue Nov 19, 2019 7:48 pm
Full name: Alayan Feh

Re: Chess variants made with help from alpha zero article

Post by Alayan » Sat Sep 12, 2020 3:13 am

mmt wrote:
Fri Sep 11, 2020 3:14 am
We don't know what effect the no-stalemate variant would have on human GM draws, though. It could be minimal. Weak players have few draws in their games so that doesn't matter much. Maybe it's possible to go through the database of highly rated human games to look at endings and try to estimate the impact. We do know that it won't be anywhere near to what armageddon scoring with no-black-castling can do and that's still not a radical change except for memorized openings.
I'd expect 10-15% more decisive games in GM play, accounting for the fact that non-fighting draws where both players have no intention whatsoever of playing for the win will remain drawn.
mmt wrote:
Fri Sep 11, 2020 3:14 am
I don't see a reason to think that Alpha Zero is risk-averse. The percentage of draws is 53% for players with Elo above 2600 in classical games (https://en.chessbase.com/post/has-the-n ... -increased).
Leela Chess Zero is notoriously risk-averse and so there is all the reasons to believe A0 is too. The giveaway is the winrate from inferior positions against weaker engines. Leela will play for the draw from the start as black instead of keeping the game alive, resulting in being worse at beating up weaker engines.
mmt wrote:
Fri Sep 11, 2020 3:14 am
Maybe I'm not up-to-date but I haven't seen any forced-opening-lines human chess tournaments and as you said, we mostly need humans to adopt it. I don't think they would - you're basically taking away a part of the game.
Forced openings are exclusive to computer chess. They don't need to be adopted by human players to be relevant for engine programmers, because any engine programmed for standard chess handles it correctly

Something like stalemate=win, armageddon, and so on, on the other hand, cause significant suboptimal play issues in engines that are not made to consider the variant.
mmt wrote:
Fri Sep 11, 2020 3:14 am
As I wrote earlier, I think forced-openings are a non-starter for humans, so we should compare to classical chess. Making computer games more interesting to watch by making every game a win instead of just half of them in the forced-opening variant is a huge difference.
When a SuFi opening goes 1.5-0.5 it's just as interesting as if armageddon scoring had been used and the game pair went 2-0 instead of 1-1 (remember, with armageddon scoring, it's either 2-0 or 1-1). If the game pair is a dull 1-1 draw in classical chess, it'll be boring. But so would be a armageddon 1-1 with both sides getting an easy win. The busted Latvian from TCEC S18 didn't bring huge excitement despite having 2 wins.

The whole argument about armageddon being exciting thus assumes that the position selected to use armageddon on is right at the limit between (normal) draw and win.
mmt wrote:
Fri Sep 11, 2020 3:14 am
1. Reverse games would require twice as much time as regular games for humans
2. People don't like draws. It's human psychology to prefer a winner/loser scenario and this theme has been true in sports with overtimes.
3. It allows black to "win" instead of just fighting to draw (assuming very good players). Makes it appear more fair and allows one-time games.
Forced openings are relevant for engine play, not for human chess, I never discussed them for human chess. Human chess doesn't have the problem of increasing hardware strength or increasing software strength. It has issues with pre-arranged draws and massive memorization of lines that, by making players play much better (and less varied) than they would if thinking by themselves in the opening/middlegame, increase the draw rate.

1. If you want to rely on only 1 game, then with armageddon, you also have to make sure that neither side gets a bigger advantage than white gets in classical chess. Over a tournament, or in a match, players get to play both white and black in similar amounts as a balancing element. But over a single game, a drawn game naturally has this balancing element as better play converges towards draws. But when removing draws, one side is going to be winning and opening theory will be very busy to find which lines lead to forced wins and which lines are lost. GMs may not be good enough to find this stuff over the board, but they are good enough to memorize 3500+ engine theory 20-30 ply-deep in relevant lines.

2. Draws can be fun if they're fighting, rare and you get overtime (the equivalent in chess would be shorter games, like was done in the recent MC tour event) to decide a winner. They're also more acceptable if the game is part of a larger tournament.

3. See 1.

4. The players will converge to their correct position in a swiss or round-robin event faster if you have some draws than if you have none. You'd need more games to keep the same reliability.

5. See the thread Laskos just posted : viewtopic.php?f=2&t=75080. He basically proved the point I made about the different scoring system being a coat of paint that doesn't change anything for engine chess. So only the different start position remains as an advantage for armageddon chess, but we already have this without it.

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

Re: Chess variants made with help from alpha zero article

Post by jp » Sat Sep 12, 2020 4:07 am

Alayan wrote:
Sat Sep 12, 2020 3:13 am
mmt wrote:
Fri Sep 11, 2020 3:14 am
2. People don't like draws. It's human psychology to prefer a winner/loser scenario and this theme has been true in sports with overtimes.
2. Draws can be fun if they're fighting, rare and you get overtime (the equivalent in chess would be shorter games, like was done in the recent MC tour event) to decide a winner. They're also more acceptable if the game is part of a larger tournament.
I'd say most deep games and sports allow the possibility of a draw, just as in war and life there isn't necessarily a winner.

People who don't accept draws in sport are probably a bit shallow.

mmt
Posts: 274
Joined: Sun Aug 25, 2019 6:33 am
Full name: .

Re: Chess variants made with help from alpha zero article

Post by mmt » Sat Sep 12, 2020 4:51 am

Alayan wrote:
Sat Sep 12, 2020 3:13 am
Leela Chess Zero is notoriously risk-averse and so there is all the reasons to believe A0 is too. The giveaway is the winrate from inferior positions against weaker engines. Leela will play for the draw from the start as black instead of keeping the game alive, resulting in being worse at beating up weaker engines.
This just comes down to the training set and to scoring. If you train against a good mix of opponents instead of just the best ones or if you score wins higher, a neural net will naturally learn to play with the equivalent of contempt.
Alayan wrote:
Sat Sep 12, 2020 3:13 am
Forced openings are relevant for engine play, not for human chess, I never discussed them for human chess.
You haven't addressed the fact that you're also taking away an important part of the game from computers. What if I have an engine that prefers some openings over others, just like most humans do? It could do well in them but not in forced openings and the assigned rating wouldn't reflect the rating it would get in normal play. Same thing if I have an engine that's better in openings in general. It's a major flaw.
Last edited by mmt on Sat Sep 12, 2020 4:54 am, edited 2 times in total.

mmt
Posts: 274
Joined: Sun Aug 25, 2019 6:33 am
Full name: .

Re: Chess variants made with help from alpha zero article

Post by mmt » Sat Sep 12, 2020 4:52 am

jp wrote:
Sat Sep 12, 2020 4:07 am
I'd say most deep games and sports allow the possibility of a draw, just as in war and life there isn't necessarily a winner.

People who don't accept draws in sport are probably a bit shallow.
The best comparison is other board games. I haven't done a survey of board games or anything but the ones I played don't have draws or they are so rare as to be unimportant.

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

Re: Chess variants made with help from alpha zero article

Post by jp » Sat Sep 12, 2020 5:15 am

mmt wrote:
Sat Sep 12, 2020 4:52 am
The best comparison is other board games. I haven't done a survey of board games or anything but the ones I played don't have draws or they are so rare as to be unimportant.
But most board games are shallow! :wink:

I'm not familiar with most board games, so let me see...

Go has draws, but "From experience the value of playing first is about 7 points, so this is the normal size of komi. In tournaments, komi is often set at 7.5 points to avoid draws." (britgo.org)

Shogi has draws, but they are very rare.

Xiangqi has draws.

Checkers has draws.

Backgammon, they say, has no draws.

Post Reply