Komodo 10.4 is the new Theban Chess Champion.

[Laskos]

Shortly:
I proposed possibly a viable game, which uses mostly the rules and pieces of standard chess, as much as chess engines like Stockfish and Komodo can play it.
(2011) http://www.talkchess.com/forum/viewtopic.php?t=40990
(2014) http://www.talkchess.com/forum/viewtopic.php?t=52277

Basically I change the starting position and the starting set of pieces, the rest being the same (almost, there are rules for first rank pawns). Many engines will refuse to play with pawns on the first rank or with more than 8 pawns, but there are _chess_ engines which do play this game, most of top engines (Houdini is a notable exception), although they are not built and tuned to play it. Also, the starting position is not legal in regular chess.

The starting position is:
[D]1p6/2p3kn/3p2pp/4pppp/5ppp/8/PPPPPPPP/PPPPPPKN w - - 0 0[/D]

Spartans are white, in their standard formation, having the King (always on the right wing in Sparta), a Knight (Hippeis), and 14 pawns, a homogenous phalanx. (Note an inaccuracy in "Spartan Chess", Spartans never had both Kings in a field battle.)

Thebans are black, in their formation of Pelopidas and Epaminondas against the Spartans. The King and the Knight (Sacred Band) are on the left flank, the 12 pawns are forming a deep column opposing the Spartan King, and the phalanx in echelon formation.

These two formations are pretty historically accurate, with Thebans in numerical inferiority (12 pawns against 14).

Rules, pieces and moves are all the same as in the standard chess (aside first rank pawns, which are illegal in regular chess).

Championship results 2014 and 2016 played in my garage were won by Stockfish. This year, the clear winner is Komodo 10.4. Also, this year, two new top engines can participate, as they now accept this variant, Deep Shredder 13 and Fire 5.

60'' + 0.6'' time control (STC):

Code: Select all

   # PLAYER                      : RATING  ERROR    POINTS  PLAYED     (%)   CFS(next)
   
   1 Komodo 10.4 64-bit          : 2724.6   34.3     319.0     400    79.8     100    
   
   2 Fire 5 x64                  : 2582.5   30.1     248.0     400    62.0      74    
   3 Stockfish 100517 64 BMI2    : 2568.3   29.2     240.0     400    60.0      97    
   4 Andscacs 0.91n              : 2527.4   28.8     216.5     400    54.1      96    
   5 Deep Shredder 13 x64        : 2488.1   29.0     193.5     400    48.4      91    
   6 Hakkapeliitta TCEC          : 2458.0   31.0     176.0     400    44.0      78    
   7 DiscoCheck 5.2.1            : 2439.8   30.1     165.5     400    41.4      89    
   8 Cheng 4.39                  : 2411.6   29.8     149.5     400    37.4     100    
   9 Quazar 0.4 x64              : 2299.7   34.5      92.0     400    23.0     ---    

White performance = 51.5%
White advantage = 13.17 +/- 8.10
Draw rate (equal opponents) = 16.31 % +/- 0.92

600'' + 6'' time control (LTC):

Code: Select all

   # PLAYER                      : RATING  ERROR    POINTS  PLAYED     (%)   CFS(next)

   1 Komodo 10.4 64-bit          : 2746.3   56.2     130.5     160    81.6     100    

   2 Stockfish 100517 64 BMI2    : 2605.5   49.6     104.0     160    65.0      80    
   3 Fire 5 x64                  : 2575.5   48.1      97.5     160    60.9      84    
   4 Andscacs 0.91n              : 2539.7   47.1      89.5     160    55.9      88    
   5 Deep Shredder 13 x64        : 2498.0   46.9      80.0     160    50.0      95    
   6 Hakkapeliitta TCEC          : 2441.0   48.4      67.0     160    41.9      89    
   7 Cheng 4.39                  : 2393.6   50.4      56.5     160    35.3      70    
   8 DiscoCheck 5.2.1            : 2374.9   47.7      52.5     160    32.8      91    
   9 Quazar 0.4 x64              : 2325.7   51.0      42.5     160    26.6     ---    

White performance = 52.2%
White advantage = 20.10 +/- 13.44
Draw rate (equal opponents) = 18.21 % +/- 1.59

Previously, in 2014 and 2016, the title was won by Stockfish. To remark good performance of Komodo, Andscacs and Hakkapeliitta. For this yest I built an opening book of several hundreds openings (I can upload the EPD file if some want it). The game is balanced White-Black, with very few Draws. The scaling with 10x increase in time control doesn't indicate that the game may become unbalanced or that the draw rate goes much higher with time control. The ratings don't compress. To very high depths Komodo 10.4 seems to give high diminishing advantage for White of 0.88. It seems Komodo 10.4 to depth 42 (about 7 hours on an i7 4 core machine) doesn't solve this game or goes to some extreme, it is not that easy for engines.

Code: Select all

info depth 40 time 14597119 nodes 123504350119 score cp 91 nps 8460871 hashfull 972 tbhits 0 pv a2a4 g7f6 b2b4 c7c5 b4b5 f6e6 c2c4 d6d5 d2d3 h7f6 a4a5 e5e4 b1b2 e4d3 e2d3 f4f3 h2h3 f6d7 e1e2 e6e5 b2b3 g4h3 g2f3 g5g4 g1h2 d5d4 f3g4 h5g4 f2f3 d7f8 a5a6 f8d7 f1f2 h6h5 a1a2 e5f4 a2a3 f4e5 a3a4 e5f4 a4a5 d7f6

info depth 41 time 18758287 nodes 159858269066 score cp 92 nps 8522007 hashfull 972 tbhits 0 pv b2b4 g7f6 b1b2 c7c5 c2c3 c5b4 c3b4 d6d5 c1c2 h7f8 c2c3 f8e6 a2a4 f6e7 d2d3 e7d6 a1a2 f4f3 e2e3 h4h3 a4a5 h5h4 e1e2 e5e4 d3d4 e6c7 a2a4 c7e8 b4b5 d6c7 b2b3 e8f6 c3c4 h3g2 f1g2 f3e2 d1e2 f5f4 e3f4 g5f4 g1f1 h4h3 g2h3 d5c4 b3c4

info depth 42 time 23883486 nodes 205241855258 score cp 88 nps 8593462 hashfull 972 tbhits 0 pv b2b4 g7f6 b1b2 c7c5 b4c5 d6c5 c2c3 f6e6 a2a4 e6d5 b2b3 h7f6 c3c4 d5d6 c1c2 f6d7 d2d3 b8b7 a1a2 d7b8 c2c3 b8c6 f2f3 g4g3 e2e4 f5e4 d3e4 g3h2 g1h2 g5g4 d1d2 g6g5 e1e2 d6e6 h2g1 g4g3 f1f2 e6d6 d2d3 g5g4 g1f1 b7b6 f2g3 h4g3 f1e1 g4f3 e2f3 h5h4

[shrapnel]

Nice, but Standard Chess is difficult enough.

[Nordlandia]

Nice, but Standard Chess is difficult enough.

Purist

Some chess variants are fun.

[cdani]

Interesting tournament! Thanks!

The "n" version of Andscacs (Andscacs 0.91n) is slower than Andscacs 0.91, as it does not have popcnt. And the b one Andscacs 0.91b is BMI2 (good for intel, bad for ryzen).

[lkaufman]

I'm surprised these engines can even play with pawns on the first rank. I have no idea why Komodo is so far above Stockfish in this variant, but probably it doesn't have much to do with ordinary chess.
I've often wondered what would be the smallest change one could make to the rules of chess to eliminate draws while making the chances of White and Black close to equal. Armageddon playoff games solve the problem by giving Black draw odds and giving White extra time, but let's assume that we want a solution that doesn't involve time. Preferably it should be one that allows current engines to play the game well. Here are some ideas:
1. Game starts with white pawn on e4, WTM (or start with moves 1.e3 Nc6 2.e4 Nb8). Black gets draw odds. Engines with simple Contempt (i.e. only based on setting value of draw, like SF I think) can play this way with no new coding (just set huge Contempt value with the right sign), others like Komodo and Houdini would need a new UCI option for drawscore to play this well.
2. White starts without c2 pawn (or start with a move sequence to discard this pawn) but Black gets draw odds. Same issues as above, but very different strategies.
3. Black cannot castle on same side as White, but gets draw odds. This would require new engine coding. A cruder version that would not require such coding is that White can only castle Kingside, Black only queenside but Black gets draw odds.
4. All draws adjudicated by some simple point count involving the ranks of the pawns and values for the pieces, with tied point counts going to Black. Current engines would play this pretty well, but of course special coding would help a lot.

I wonder what ideas others might suggest.

[JJJ]

Maybe Komodo is better than Stockfish with some pieces than others on the board. Maybe Stockfish is better than Komodo with Rook, Bishop and less good with Queen and Knight ?

[Laskos]

I'm surprised these engines can even play with pawns on the first rank. I have no idea why Komodo is so far above Stockfish in this variant, but probably it doesn't have much to do with ordinary chess.
I've often wondered what would be the smallest change one could make to the rules of chess to eliminate draws while making the chances of White and Black close to equal. Armageddon playoff games solve the problem by giving Black draw odds and giving White extra time, but let's assume that we want a solution that doesn't involve time. Preferably it should be one that allows current engines to play the game well. Here are some ideas:
1. Game starts with white pawn on e4, WTM (or start with moves 1.e3 Nc6 2.e4 Nb8). Black gets draw odds. Engines with simple Contempt (i.e. only based on setting value of draw, like SF I think) can play this way with no new coding (just set huge Contempt value with the right sign), others like Komodo and Houdini would need a new UCI option for drawscore to play this well.
2. White starts without c2 pawn (or start with a move sequence to discard this pawn) but Black gets draw odds. Same issues as above, but very different strategies.
3. Black cannot castle on same side as White, but gets draw odds. This would require new engine coding. A cruder version that would not require such coding is that White can only castle Kingside, Black only queenside but Black gets draw odds.
4. All draws adjudicated by some simple point count involving the ranks of the pawns and values for the pieces, with tied point counts going to Black. Current engines would play this pretty well, but of course special coding would help a lot.

I wonder what ideas others might suggest.

I tried 1., but Black is heavily favored, because many Draws occur. Score was in 200 games 32% White, 68% Black (with Draws scored as Black Wins). In 2. you mean White gets Draw odds, right? 3. probably will favor heavily Black again, because Draws are only mildly eliminated. 4. would work the best probably. I would love to test it, but special Win rules must be put and coded into engines.

[Laskos]

Interesting tournament! Thanks!

The "n" version of Andscacs (Andscacs 0.91n) is slower than Andscacs 0.91, as it does not have popcnt. And the b one Andscacs 0.91b is BMI2 (good for intel, bad for ryzen).

Thanks, I replayed STC games, the results are similar. Very good performance by Andscacs!

Code: Select all

   # PLAYER                      : RATING  ERROR    POINTS  PLAYED     (%)   CFS(next)
   1 Komodo 10.4 64-bit          : 2720.5   35.4     317.5     400    79.4     100    
   2 Stockfish 100517 64 BMI2    : 2580.8   31.5     247.5     400    61.9      53    
   3 Fire 5 x64                  : 2579.0   30.0     246.5     400    61.6      97    
   4 Andscacs 0.91b              : 2536.1   29.3     222.0     400    55.5      84    
   5 Deep Shredder 13 x64        : 2513.8   29.0     209.0     400    52.3     100    
   6 Hakkapeliitta TCEC          : 2419.9   30.8     154.5     400    38.6      74    
   7 DiscoCheck 5.2.1            : 2404.7   30.3     146.0     400    36.5      80    
   8 Cheng 4.39                  : 2385.7   31.4     135.5     400    33.9      86    
   9 Quazar 0.4 x64              : 2359.5   31.4     121.5     400    30.4     ---    

White advantage = -7.74 +/- 8.40
Draw rate (equal opponents) = 15.94 % +/- 0.89

[cdani]

Interesting tournament! Thanks!

The "n" version of Andscacs (Andscacs 0.91n) is slower than Andscacs 0.91, as it does not have popcnt. And the b one Andscacs 0.91b is BMI2 (good for intel, bad for ryzen).

Thanks, I replayed STC games, the results are similar. Very good performance by Andscacs!

Thanks! But was not necessary I just told it for other occasions.

[lkaufman]

I'm surprised these engines can even play with pawns on the first rank. I have no idea why Komodo is so far above Stockfish in this variant, but probably it doesn't have much to do with ordinary chess.
I've often wondered what would be the smallest change one could make to the rules of chess to eliminate draws while making the chances of White and Black close to equal. Armageddon playoff games solve the problem by giving Black draw odds and giving White extra time, but let's assume that we want a solution that doesn't involve time. Preferably it should be one that allows current engines to play the game well. Here are some ideas:
1. Game starts with white pawn on e4, WTM (or start with moves 1.e3 Nc6 2.e4 Nb8). Black gets draw odds. Engines with simple Contempt (i.e. only based on setting value of draw, like SF I think) can play this way with no new coding (just set huge Contempt value with the right sign), others like Komodo and Houdini would need a new UCI option for drawscore to play this well.
2. White starts without c2 pawn (or start with a move sequence to discard this pawn) but Black gets draw odds. Same issues as above, but very different strategies.
3. Black cannot castle on same side as White, but gets draw odds. This would require new engine coding. A cruder version that would not require such coding is that White can only castle Kingside, Black only queenside but Black gets draw odds.
4. All draws adjudicated by some simple point count involving the ranks of the pawns and values for the pieces, with tied point counts going to Black. Current engines would play this pretty well, but of course special coding would help a lot.

I wonder what ideas others might suggest.

I tried 1., but Black is heavily favored, because many Draws occur. Score was in 200 games 32% White, 68% Black (with Draws scored as Black Wins). In 2. you mean White gets Draw odds, right? 3. probably will favor heavily Black again, because Draws are only mildly eliminated. 4. would work the best probably. I would love to test it, but special Win rules must be put and coded into engines.

1. Surprising, as the eval is close to what we think is the win/draw threshold. This could tell us something.
2. Yes, of course I meant White gets Draw odds.
3. The point here is that it is usually rather difficult for Black to castle long in most openings; Black would have to play some passive space-conceding setup. Maybe still now enough for draw odds, but should at least be close.
4. White probably gets a small space advanage by force in the opening, so the rules might have to say something like "advancement difference of less than 2 is a win for Black". It's like Komi in Go; the second player wins unless the first player has 7 (or 6 or 8) more points of space than the second player.