There have been many draws in division 2 so far, the engine scores range between 3 and 6.5 after 10 rounds. ChessBrain and Ethereal are the current leaders, Texel, Arasan and Xiphos are not far behind. Ethereal is the only engine without loss so far and still anything can happen.
We follow the ChessBrain - Texel game in round 11. The reverse game ended in a repetition after 31 moves.
move 6: Texel diverged from the reverse game on its first move. ChessBrain's 5th move was less popular, its eval is climbing though, at 0.5 after move 6.
move 11: The engines exchanged DS bishops, ChessBrain's eval continues to increase and Texel's eval remains close to 0. Both kings castled and the position looks quiet. ChessBrain isn't showing its plan, its PV is usually empty.
move 15: Texel is advancing pawns on the queen side, ChessBrain may be preparing a king side attack. The black LS bishop is still looking for a good square, b7 and d7 are its only options so far and they're not great. After exchanging knights in the center ChessBrain has a pawn on e5 and the black king side looks vulnerable.
move 19: Texel's eval started to climb while ChessBrain's came down a little and then jumped over 1 as the white knight landed on d4. ChessBrain finally shared a PV with us, I'm not sure the ending it sees is winning but in any case Texel is not following the line.
move 25: ChessBrain doesn't want to exchange bishops, hides its bishop on b1 still looking at the black king side. With the queen on the same diagonal finally there is a real threat on the black king, forcing pawn moves and weakening the black king side.
move 30: There is some pawn action on the queen side, perhaps Texel would like to open files for its rooks. Evals seem to be cooling down a little.
move 35: Texel is attacking the c3 pawn and it seems that white lost its momentum. Texel opened the a file and ChessBrain thinks this is good, its eval over 1 again. Now Texel attacks the king side with g4.
move 40: There are two isolated black pawns on h4 and b5, ChessBrain can try to capture them. ChessBrain eval jumps close to 2 and Texel's eval over 1, can ChessBrain covert here?
move 48: ChessBrain is a pawn up with a queen side passer. Evals are climbing fast, ChessBrain moves its pieces to the king side and captures the h pawn. Texel exchanged knight for bishop but I don't think it can hold two pawns down.
move 55: A pair of rooks is exchanged, only QRN vs QRB left. White can attack with its pawns on both sides, the win rule will stop the game shortly.
move 58: Game over, well played ChessBrain.
ChessBrain is in the lead with +4 after 11 games, still many games to play until the end of the division.
Sunday, August 19, 2018
Saturday, August 18, 2018
Season 13 division 3 statistics
Draw rate, wins
Final draw rate was 42.9%, relatively low.
Game termination
The three most common game termination causes were:
49.1% - TCEC win rule
21.4% - SyzygyTB
17.0% - TCEC draw rule
There were 3 crashes in the division, and two games had to be replayed due to a server connection problem.
Moves per game
Median= 57
Average= 66.5
There were 15 games longer than 100 moves, the longest was 176 moves (Leela - Ethereal, round 24, draw).
Time per game (hours)
Median= 1:10
Average= 1:12
Openings
There were 2-move book openings in this stage chosen by Cato. Each opening was repeated 4 times in a round, and then again in the reverse games round. In total 14 2-move openings were used. The first letter of the ECO codes was distributed as follows:
If we use the opening 'family name' (using format FAMILY_NAME: VARIANT....) the top 3 are:
Sicillian - 23 times
Queen's pawn game - 9 times
English, Nimzovich-Larsen attack, Petrov - 8 times
The engines had a lot of freedom to choose the opening variant. However, 48.2% of the game pairs repeated the same ECO code twice, and 32.1% repeated the same opening variant. In 19.6% of the game pairs the ECO first letter was not repeated twice.
Reverse pairs, wins
Reverse pairs, same moves
Pairs of reverse games diverged very quickly, 82.1% of the pairs diverged at most after 1 move. The longest repeated move sequence in a game pair was 14 plys long, in DeusX-Hannibal, rounds 7 and 14, both draws.
Final draw rate was 42.9%, relatively low.
Game termination
The three most common game termination causes were:
49.1% - TCEC win rule
21.4% - SyzygyTB
17.0% - TCEC draw rule
There were 3 crashes in the division, and two games had to be replayed due to a server connection problem.
Moves per game
Median= 57
Average= 66.5
There were 15 games longer than 100 moves, the longest was 176 moves (Leela - Ethereal, round 24, draw).
Time per game (hours)
Median= 1:10
Average= 1:12
Openings
There were 2-move book openings in this stage chosen by Cato. Each opening was repeated 4 times in a round, and then again in the reverse games round. In total 14 2-move openings were used. The first letter of the ECO codes was distributed as follows:
If we use the opening 'family name' (using format FAMILY_NAME: VARIANT....) the top 3 are:
Sicillian - 23 times
Queen's pawn game - 9 times
English, Nimzovich-Larsen attack, Petrov - 8 times
The engines had a lot of freedom to choose the opening variant. However, 48.2% of the game pairs repeated the same ECO code twice, and 32.1% repeated the same opening variant. In 19.6% of the game pairs the ECO first letter was not repeated twice.
Reverse pairs, wins
Reverse pairs, same moves
Pairs of reverse games diverged very quickly, 82.1% of the pairs diverged at most after 1 move. The longest repeated move sequence in a game pair was 14 plys long, in DeusX-Hannibal, rounds 7 and 14, both draws.
Friday, August 17, 2018
Season 13 division 3 final results
Division 3 final crosstable
Ethereal is promoted to division 2 and Bobcat is relegated to division 4. Places 2 and 7 in the table were decided by tiebreak. Arasan beat Leela in their head-to-head matches (one win in 4 games, 2.5-1.5) and it is the second engine to promote. Nemorino crashed twice while Hannibal crashed once, so Nemorino is the second engine relegated.
After RR3 the scores were Arasan +3, Leela +1, Pedone 0, DeusX 0, Hannibal -2, Nemorino -3. Leela's first game in RR4 against Hannibal ended in a Leela disconnect, this was later determined to be a server problem and the game had to be replayed. After 3 rounds were played and the crash result fixed the scores were Arasan +5, Leela +3, Pedone 0, DeusX -1, Nemorino -2, Hannibal -5. Leela and Arasan drew their game, two rounds to go the scores were Arasan +4, Leela +4, Pedone 0, DeusX 0, Nemorino -3, Hannibal -5. In round 27 Arasan lost to Pedone, Leela held a draw against DeusX, Hannibal and Nemorino also drew their game. The situation before the last round was Leela +4, Arasan +3, Nemorino -3, Hannibal -5. Pedone forced a perpetual check draw against Leela, then Arasan beat Bobcat equalizing points with Leela and reaching 2nd place. In the last two games Nemorino lost to Ethereal and Hannibal surprised DeusX, closing the gap in the last game.
Both neural network engines remain in division 3. There were several hardware issues, the GPU configuration was changed for division 3 and then changed again while division 3 was played. Connection problems caused games to crash and they had to be replayed. The many Leela fans were disappointed by the result, claiming that under good GPU conditions Leela would have promoted easily. Considering that Arasan got a point from a Nemorino crash in a lost position Leela fans have another reason to believe that Leela could have promoted. It is probable that in the next TCEC season the hardware will improve and Leela will continue to move up in the divisions.
Division 2 will start later today with the engines Gull, Texel, Xiphos, ChessBrainVB, Nirvana, Vajolet, Ethereal and Arasan. Time control will continue to be 30+10. After dominating division 3 Ethereal is trying to promote again, we will see if division 2 engines are tougher opponents.
Ethereal is promoted to division 2 and Bobcat is relegated to division 4. Places 2 and 7 in the table were decided by tiebreak. Arasan beat Leela in their head-to-head matches (one win in 4 games, 2.5-1.5) and it is the second engine to promote. Nemorino crashed twice while Hannibal crashed once, so Nemorino is the second engine relegated.
After RR3 the scores were Arasan +3, Leela +1, Pedone 0, DeusX 0, Hannibal -2, Nemorino -3. Leela's first game in RR4 against Hannibal ended in a Leela disconnect, this was later determined to be a server problem and the game had to be replayed. After 3 rounds were played and the crash result fixed the scores were Arasan +5, Leela +3, Pedone 0, DeusX -1, Nemorino -2, Hannibal -5. Leela and Arasan drew their game, two rounds to go the scores were Arasan +4, Leela +4, Pedone 0, DeusX 0, Nemorino -3, Hannibal -5. In round 27 Arasan lost to Pedone, Leela held a draw against DeusX, Hannibal and Nemorino also drew their game. The situation before the last round was Leela +4, Arasan +3, Nemorino -3, Hannibal -5. Pedone forced a perpetual check draw against Leela, then Arasan beat Bobcat equalizing points with Leela and reaching 2nd place. In the last two games Nemorino lost to Ethereal and Hannibal surprised DeusX, closing the gap in the last game.
Both neural network engines remain in division 3. There were several hardware issues, the GPU configuration was changed for division 3 and then changed again while division 3 was played. Connection problems caused games to crash and they had to be replayed. The many Leela fans were disappointed by the result, claiming that under good GPU conditions Leela would have promoted easily. Considering that Arasan got a point from a Nemorino crash in a lost position Leela fans have another reason to believe that Leela could have promoted. It is probable that in the next TCEC season the hardware will improve and Leela will continue to move up in the divisions.
Division 2 will start later today with the engines Gull, Texel, Xiphos, ChessBrainVB, Nirvana, Vajolet, Ethereal and Arasan. Time control will continue to be 30+10. After dominating division 3 Ethereal is trying to promote again, we will see if division 2 engines are tougher opponents.
Wednesday, August 15, 2018
Season 13 live blogging, Wednesday Aug 15th
Division 3 is in the 3rd RR out of 4. Ethereal is leading without loss and at least 3 points away from all the other engines. There are 3 engines in the race for second place. Pedone has 11 points, DeusX has 11 and one game less, and Arasan has 10.5 points. Leela with 9 points needs to win games to have a chance, so far it has only 2 wins in 19 games.
Currently the Nemorino - DeusX game is running, closing the 19th round.
move 15: It seems that DeusX went wrong very early in this opening. It did not follow the usual line in the Petrov opening, and Nemorino's eval was over 1.5 from move 6. DeusX lost a knight for a pawn on move 9, and Nemorino's eval was over 2 on move 14.
move 20: Evals are not moving, DeusX is trying to hold. Nemorino has a queen side passer and has a potential pawn attack on the king side since the engines castled in opposite directions.
move 25: The white knight is in a weak square, not trapped but doesn't have a good square to go to. DeusX is attacking the queen side, evals are increasing again.
move 30: After several pawn exchanges it seemed DeusX would give a rook for a bishop but it changed its mind. Nemorino's king ran to the king side, it is a knight and pawn up but its king side rook is stuck in the corner.
move 32: BOOM, evals over 5. DeusX seems to have missed some tactic, and it immediately realized it is losing. It seems the black king is under immediate threat and black is about to lose material.
move 37: Game over, adjudicated by the win rule. DeusX hasn't lost material yet but its king is under attack and only the queen is there to help. Both engines agreed black's position is hopeless.
With this loss DeusX remains 3rd 0.5 point behind Pedone and equal to Arasan. The race is going to be close to the end of the division.
Currently the Nemorino - DeusX game is running, closing the 19th round.
move 15: It seems that DeusX went wrong very early in this opening. It did not follow the usual line in the Petrov opening, and Nemorino's eval was over 1.5 from move 6. DeusX lost a knight for a pawn on move 9, and Nemorino's eval was over 2 on move 14.
move 20: Evals are not moving, DeusX is trying to hold. Nemorino has a queen side passer and has a potential pawn attack on the king side since the engines castled in opposite directions.
move 25: The white knight is in a weak square, not trapped but doesn't have a good square to go to. DeusX is attacking the queen side, evals are increasing again.
move 30: After several pawn exchanges it seemed DeusX would give a rook for a bishop but it changed its mind. Nemorino's king ran to the king side, it is a knight and pawn up but its king side rook is stuck in the corner.
move 32: BOOM, evals over 5. DeusX seems to have missed some tactic, and it immediately realized it is losing. It seems the black king is under immediate threat and black is about to lose material.
move 37: Game over, adjudicated by the win rule. DeusX hasn't lost material yet but its king is under attack and only the queen is there to help. Both engines agreed black's position is hopeless.
With this loss DeusX remains 3rd 0.5 point behind Pedone and equal to Arasan. The race is going to be close to the end of the division.
Tuesday, August 14, 2018
Season 13 live blogging, Monday Aug 13th
Division 3 is getting close to the half way point. So far Ethereal is leading with a remarkable result of 8.5/10. Bobcat and Hannibal seem to be heading for relegation, and the remaining 5 engines - including the two NN engines Leela and DeusX - are in the struggle for the second promotion spot. Nemorino has had two early crashes and has been restricted to 16 threads, so far avoiding the 3rd strike and still playing well enough to have a chance of promotion. Pedone and Arasan are the remaining two engines in the race.
We will look at the Arasan-Ethereal game of round 11, Ethereal won the reverse game in round 4.
move 4: diverging from the reverse game right out of book, with a rare move e5. Arasan started optimistic with an eval around 0.5.
move 7: Ethereal eval dropped to 0 quickly. Arasan chose a rare opening line, Ethereal's eval turned slightly negative. Both the white knights are pinned.
move 10: The queen side is open, black has not developed the king side yet, delaying its castling plans. Ethereal gets a pawn, eval back to 0.
move 13: Ethereal with a very long think, but eval still 0. There is no castling in its PV, it created a pawn line across the board and it expects a lot of shuffling.
move 21: Evals are not moving, Ethereal is a pawn ahead, its king has not castled though it now can, Arasan has a bishop pair, both queens have been pushed back. Both engines seem to be playing defensively, unless there is some breakthrough I think the game will continue with a long shuffle and a draw.
move 24: Arasan is planning to trade bishops, another long think from Ethereal, again eval stays at 0.
move 33: All bishops are gone. Arasan got its pawn back and now its eval is also at 0. Perhaps the game will end at move 40 with the draw rule?
move 40: Ethereal's eval started to jump, reaching negative 0.6. Is the white back rank vulnerable? After queens are exchanged only RN vs RN left. Both evals are negative but not high, still drawish?
move 50: The queen side pawns are gone and evals are not going anywhere. Still looks like a draw, waiting for the engines to lower the eval so the game will be adjudicated.
move 60: The engines finally dropped their eval to 0. Is the end near? The 50-move counter is at 34.
move 69: Finally adjudicated by the draw rule. Arasan holds the reverse game, neither engine had a significant advantage in the game. Ethereal still unbeaten in the division, Arasan still strong in the promotion race.
We will look at the Arasan-Ethereal game of round 11, Ethereal won the reverse game in round 4.
move 4: diverging from the reverse game right out of book, with a rare move e5. Arasan started optimistic with an eval around 0.5.
move 7: Ethereal eval dropped to 0 quickly. Arasan chose a rare opening line, Ethereal's eval turned slightly negative. Both the white knights are pinned.
move 10: The queen side is open, black has not developed the king side yet, delaying its castling plans. Ethereal gets a pawn, eval back to 0.
move 13: Ethereal with a very long think, but eval still 0. There is no castling in its PV, it created a pawn line across the board and it expects a lot of shuffling.
move 21: Evals are not moving, Ethereal is a pawn ahead, its king has not castled though it now can, Arasan has a bishop pair, both queens have been pushed back. Both engines seem to be playing defensively, unless there is some breakthrough I think the game will continue with a long shuffle and a draw.
move 24: Arasan is planning to trade bishops, another long think from Ethereal, again eval stays at 0.
move 33: All bishops are gone. Arasan got its pawn back and now its eval is also at 0. Perhaps the game will end at move 40 with the draw rule?
move 40: Ethereal's eval started to jump, reaching negative 0.6. Is the white back rank vulnerable? After queens are exchanged only RN vs RN left. Both evals are negative but not high, still drawish?
move 50: The queen side pawns are gone and evals are not going anywhere. Still looks like a draw, waiting for the engines to lower the eval so the game will be adjudicated.
move 60: The engines finally dropped their eval to 0. Is the end near? The 50-move counter is at 34.
move 69: Finally adjudicated by the draw rule. Arasan holds the reverse game, neither engine had a significant advantage in the game. Ethereal still unbeaten in the division, Arasan still strong in the promotion race.
Saturday, August 11, 2018
Season 13 division 4 statistics
Draw rate, wins
Final draw rate was 51.8%.
Game termination
The three most common game termination causes were:
44.6% - TCEC win rule
21.4% - TCEC draw rule
16.1% - 3-Fold repetition
There were no crashes in the division.
Moves per game
Median= 68.5
Average= 76.7
The games were unusually long compared to previous seasons. Without testing this I have a feeling that the NN engine games were longer on average, as a result of their different attitude towards evaluation. When winning their opponent sometimes was not aware it was losing for a while, and when drawing the NN engines sometimes took more time to realize the game was a draw.
There were 23 games longer than 100 moves, the longest was 272 moves (Rodent - DeusX, round 27, draw).
Time per game (hours)
Median= 1:17
Average= 1:17
Openings
There were 2-move book openings in this stage chosen by Cato. The first letter of the ECO codes was distributed as follows:
If we use the opening 'family name' (using format FAMILY_NAME: VARIANT....) the top 3 are:
Sicillian - 18 times
QGD - 17 times
English - 10 times
The engines had a lot of freedom to choose the opening variant. Only 19.6% of the game pairs repeated the same ECO code twice, and only 14.3% repeated the same opening variant. In 12.5% of the game pairs the ECO first letter was not repeated twice.
Reverse pairs, wins
Reverse pairs, same moves
Pairs of reverse games diverged very quickly, 55.3% diverged immediately out of book, 87.5% of the pairs diverged at most after 1 move.
Final draw rate was 51.8%.
Game termination
The three most common game termination causes were:
44.6% - TCEC win rule
21.4% - TCEC draw rule
16.1% - 3-Fold repetition
There were no crashes in the division.
Moves per game
Median= 68.5
Average= 76.7
The games were unusually long compared to previous seasons. Without testing this I have a feeling that the NN engine games were longer on average, as a result of their different attitude towards evaluation. When winning their opponent sometimes was not aware it was losing for a while, and when drawing the NN engines sometimes took more time to realize the game was a draw.
There were 23 games longer than 100 moves, the longest was 272 moves (Rodent - DeusX, round 27, draw).
Time per game (hours)
Median= 1:17
Average= 1:17
Openings
There were 2-move book openings in this stage chosen by Cato. The first letter of the ECO codes was distributed as follows:
If we use the opening 'family name' (using format FAMILY_NAME: VARIANT....) the top 3 are:
Sicillian - 18 times
QGD - 17 times
English - 10 times
The engines had a lot of freedom to choose the opening variant. Only 19.6% of the game pairs repeated the same ECO code twice, and only 14.3% repeated the same opening variant. In 12.5% of the game pairs the ECO first letter was not repeated twice.
Reverse pairs, wins
Reverse pairs, same moves
Pairs of reverse games diverged very quickly, 55.3% diverged immediately out of book, 87.5% of the pairs diverged at most after 1 move.
Friday, August 10, 2018
Season 13 division 4 final results
Division 4 final crosstable
The two NN engines Leela (=LCZero) and DeusX promote to division 3. Wasp was in the race until the very last round. It managed to hold a draw against Leela, but DeusX's win against Senpai meant Wasp finished third, half a point behind DeusX.
The NN engines attracted a lot of attention, Leela in particular has a lot of supporters. One such supporter is the well-known kingscrusher, who offered live commentary of Leela games in the chat and posted analysis videos of these games on his YouTube channel. The chat was crowded and it was hard to keep up with all the posts while watching the game.
A lot of chat posts referred to the DeusX controversy. DeusX is a neural network engine that used the Leela code but used a different learning method. While Leela started from scratch and learned only playing itself, DeusX used human games to train its network. Apart from giving credit to the Leela development team for its code, there was a question whether DeusX should be considered a separate entity, or is it comparable to a Stockfish clone. TCEC decided to let both play in division 4, and there was a continuous argument about this decision in the chat. Since both engines will continue to play in division 3 I suppose the argument will continue.
There was a new TCEC site administrator during this division, indoda_eziphukile. The GUI reverted to the older version with a few changes (I will have to get used to new colors for the board). There were no problems with the system, and no crashes despite the complexity of running two machines that sometimes have to communicate. The archive did not work for the first few days but was fixed later. Well done indoda! We were then informed that for division 3 the site administrator will be replaced yet again. I hope TCEC is looking for a more stable solution.
Division 3 will start soon, with Hannibal, Ethereal, Arasan, Pedone, Bobcat, Nemorino, Leela and DeusX. The NN fans are confident Leela will promote again, it will play with a new network version which is stronger. However, considering that Wasp played in division 3 and was relegated in season 12 I think it is quite possible that division 3 will be too strong for Leela. Another issue to look out for is crashes. Arasan and Hannibal crashed out of division 2 in season 12, I hope their developers found the problems and fixed them for this season.
The two NN engines Leela (=LCZero) and DeusX promote to division 3. Wasp was in the race until the very last round. It managed to hold a draw against Leela, but DeusX's win against Senpai meant Wasp finished third, half a point behind DeusX.
The NN engines attracted a lot of attention, Leela in particular has a lot of supporters. One such supporter is the well-known kingscrusher, who offered live commentary of Leela games in the chat and posted analysis videos of these games on his YouTube channel. The chat was crowded and it was hard to keep up with all the posts while watching the game.
A lot of chat posts referred to the DeusX controversy. DeusX is a neural network engine that used the Leela code but used a different learning method. While Leela started from scratch and learned only playing itself, DeusX used human games to train its network. Apart from giving credit to the Leela development team for its code, there was a question whether DeusX should be considered a separate entity, or is it comparable to a Stockfish clone. TCEC decided to let both play in division 4, and there was a continuous argument about this decision in the chat. Since both engines will continue to play in division 3 I suppose the argument will continue.
There was a new TCEC site administrator during this division, indoda_eziphukile. The GUI reverted to the older version with a few changes (I will have to get used to new colors for the board). There were no problems with the system, and no crashes despite the complexity of running two machines that sometimes have to communicate. The archive did not work for the first few days but was fixed later. Well done indoda! We were then informed that for division 3 the site administrator will be replaced yet again. I hope TCEC is looking for a more stable solution.
Division 3 will start soon, with Hannibal, Ethereal, Arasan, Pedone, Bobcat, Nemorino, Leela and DeusX. The NN fans are confident Leela will promote again, it will play with a new network version which is stronger. However, considering that Wasp played in division 3 and was relegated in season 12 I think it is quite possible that division 3 will be too strong for Leela. Another issue to look out for is crashes. Arasan and Hannibal crashed out of division 2 in season 12, I hope their developers found the problems and fixed them for this season.
Subscribe to:
Posts (Atom)