Tuesday, September 29, 2015

Season 8 stage 2 after 14 rounds

The stage 2 crosstable after 14 rounds is:

The top 3 look solid. Houdini has had some problems earlier but now looks as a sure third place. The rest of the qualification is still very close with engines ranked 4-10 fighting for 3 places, and even Chiron still has a chance. With only 8 games to go the small advantages start to make a difference. Ginkgo and Hannibal improved their position with wins against contenders Texel and Nirvana.



Stockfish had two draws in rounds 12-14 and is now in second place.

Komodo got into a technical RNN vs R and two pawn advantage ending against Hannibal. It used its tablebase advantage to see the win long before Hannibal understood what was going on.

The Baron managed to hold out against Houdini for over 130 moves. The game reached a 7 piece tablebase draw of KRBN vs KRB, but The Baron couldn't find the right moves and lost the ending.


Notable game
Nirvana - Hannibal, round 14
Game on TCEC archive

This was an important game for both engines because of the qualification race. Nirvana had a very good first half, but was only a half-point better than Hannibal before the game. After 18 moves the position was:



The material was balanced and both evals were drawish. Hannibal had some pressure on the king side but it looked under control. Nirvana did not see the danger, it moved on the queen side and accepted a bishop sacrifice, and in a few moves was under a full attack on the king side.



With all the black pieces joining the attack Nirvana had to give back the bishop and a pawn, and the white king had to run to the center as fast as possible.



In an attempt to relieve the pressure Nirvana gave a bishop for two pawns.



This was the beginning of the end for Nirvana. Once queens and a pair of rooks were exchanged white's position crumbled and it started losing pawns. The game ended with Hannibal having a clear material advantage.

Saturday, September 26, 2015

Season 8 stage 2 stats after first round robin

These are the rankings at the halfway point:

Houdini with two wins improved its position and is back in third. Nirvana lost two games and is back in the difficult qualification struggle. With only one point separating the engines in ranks 5-10 qualifying is going to be tough.

Fire managed to reach a queen and pawns endgame against Gull, it had a pawn advantage and according to the 7-piece tablebase it was a win. However it could not convert the tricky win and could only manage a draw after 50 futile moves. Fire still has no win in this stage, but it still has a shot at qualifying.

The main attraction of rounds 10-11 was undoubtfully the Komodo-Stockfish game. This was the first of probably more that 100 games these two top engines will play this season. It had almost no implication for the qualification of this stage but everyone wanted to see the outcome, especially if one engine won. The game was very tense for the first 25 moves, mainly because Komodo kept an eval of more than 0.3 and any small jump upwards could mean the beginning of the end for Stockfish. In the end Stockfish held the game and Komodo's eval got back to zero, and at least the draw was rather quick.

Instead of the usual notable game I'll finish this post with the halfway stats.

Draw rate, wins

 The draw rate is now at 60.6%, still high but much better than in the first rounds of the stage.

Moves per game

Average = 80.48
Median = 71.25

That's a strange looking distribution. The games are longer in this stage compared to stage 1. The distance between the average and median is the result of having extra long games

Time per game (hours)

Average = 4:47
Median = 4:48

The games are over one hour longer than stage 1, reasonable since each engine has 30 more minutes to think and the games have more moves.

Openings 
Since the openings are pre-selected for the first two moves there is no point looking at the first move stats.

Using ECO codes there were 48 different openings. The top 3 ECO codes were:
A01 3 times - Nimzovich-Larsen attack
A30 3 times - English, symmetrical variation
B22 3 times - Sicillian, Alappin's variation

Using full opening name there were 63 different openings, almost no repetition so the work of organizers was a success. There were three openings that were repeated twice:
- English: Closed
- Neo-Gruenfeld, 6.O-O c6 7.Nbd2 Bf5
- Queen's Pawn: Modern

Using only the first letter of the ECO codes we get the following distribution:

The 'A's have it in this stage.

If we use the opening 'family name' (using format FAMILY_NAME: VARIANT....) the top 3 are:
English - 14 times
Sicilian - 11 times
QGD - 5 times


Wednesday, September 23, 2015

Season 8 stage 2 after 9 rounds

The crosstable after 9 rounds in stage 2 is:


Stockfish and Komodo are standing out at the top, The Baron is clearly last, but the middle is still very tense. The reverse rematches will start in two rounds, and then we will see whether there were serious biases in the openings. I have a feeling the 2-move books were balanced enough and that the table is a good representation of the engines' real strength. The draw rate has improved a little and it is now 65%.


Komodo impressed with three straight wins, it has caught up with Stockfish. Their first head to head this season will be in round 11 and will probably break the chat viewing record, although it probably will not affect their qualification for the next round.

Houdini survived its game against Stockfish, but it lost to Ginkgo after having a steady advantage and could have had an easy draw. In one of the big surprises of the stage Ginkgo managed to slowly convert a closed game into a win, applying pressure with a passed pawn on the queen side and at some point trapping Houdini's queen on the opposite corner. Houdini is now 5th and will have to work hard to qualify. Is this the first season where more engines manage to overtake Houdini? With a few more losses it may not qualify for next stage.

Nirvana decided to finish its drawn rook endgame against Chiron in style instead of waiting for the 50 move rule. It allowed Chiron to queen a pawn for a rook, and then built a fortress in a Q vs RPP endgame. Chiron was sure it was winning for 40 moves until it realized the game was a draw.

Notable game
Komodo - Ginkgo, round 9
Game on TCEC archive

After beating Houdini there was a possibility that Ginkgo could at least hold against 'the lizard', but it was not to be. A few moves into the game Komodo had a small advantage with a pawn majority on the queen side.



A few moves later black's queen side pawns are almost gone.



The white knight was on the way to c6 forcing black's pieces to move away. On the king side the black rook and bishop were too far and could not join the defense without wasting many moves. Queens were exchange and Komodo gave a rook for a bishop to gain strong connected passers on the queen side.



The white king raced to join the action, while black could only wait since its rooks could not penetrate the white diagonal wall.



Ginkgo was totally lost. The white pawns advanced with the king behind them, and the game was stopped before black lost both its rooks to stop the queenings.

Sunday, September 20, 2015

Season 8 stage 2 after 6 rounds

The stage 2 crosstable after 6 rounds looks like this:




After half of the first RR over, the highest ranked engines Stockfish, Komodo and Houdini can be found in the top half of the table. Nirvana is surprising in third place, but with so many games left anything can happen. The draw rate so far is 69%, and this will make the fight for qualifying very close and tense. Time will tell as always.


Komodo got its first wins in the rounds 4-6. The first was an easy win against The Baron. The second win was against Houdini in one of the highlights of this stage.

There have been several annoying long draws. Gull - Ginkgo, The Baron - Ginkgo, Fire - Komodo, Hannibal - Texel are examples from rounds 4-6. The game is over after 30-40 moves, both engines start shuffling pieces but at least one can't let go of some technical eval advantage that isn't moving and isn't expected to change. And when the 50-move draw gets near an engine decides it is better to move a pawn and lose some of the advantage than to end the game. ANNOYING. Only thing to do is stop watching for a few hours until they go away and a new game starts.

Nirvana - Texel looked like one of these boring long draws, but it suddenly changed after 160 moves. It's not clear whether Texel blundered, was it time pressure or could Nirvana have won a lot earlier if it had the guts to sacrifice a knight before.

Protector built an eval advantage of 1.8 over Houdini but could not turn the advantage into a win.


Stockfish - Hannibal reached a 7-piece endgame which was a tablebase draw. Hannibal didn't blunder (unlike Houdini in the game in stage 1b) and held the draw.

Notable game
Stockfish - Gull, round 6
Game on TCEC archive

After 32 moves the position was:



Both engines saw the game was heading for a draw, but Stockfish decided at least it wouldn't be boring. It started with clearing the queen side, trading 3 pieces for the black queen in the process.



The king side was open for Gull's attacking, and white traded a further rook for one of the bishops to break the attack.



This wasn't a desperate idea since white regained the rook immediately with this beautiful double rook fork by the queen



The game soon ended with a perpetual check by black. Now that is a fighting draw.

Thursday, September 17, 2015

Season 8 stage 2 after 3 rounds

Stage 2 is here!!

No weak engines left (except The Baron) and games should be harder and more balanced. Changes for this stage:
  • The openings for this stage have a preset of 2 moves to create more variability as well as excitement. 
  • All games will be played in pairs where the engines play both sides.
  • The time controls are 120 minutes per engine and 30 seconds increment per move. There is more time to think than in stage 1, and the games will be about a hour longer.
The standings after 3 rounds are:


The high draw rate came as a surprise to me. It was expected to rise compared to stage 1 but over 70% seems a bit extreme. Are the engines better? Is it the increase in time? Are the openings drawish? Will the reverse games make a difference?

Houdini and Stockfish are leading with 2.5 points, and because there were so many draws most engines have 1 or 1.5 points. Komodo has had three draws so far. In the game against Hannibal it had to fight for a draw with a rook against bishop and pawns. I expect it will have better performances in later rounds.

The Baron is playing better than expected, with two draws and a loss. It almost drew with Houdini in a game with over 150 moves containing long stretches of shuffling. The Houdini win came as a surprise.

The 2-move 'book' openings are a bit strange. Many of them are unorthodox but instead of creating unbalanced exciting games they seem to invite cautios play and drawn positions. Surprisingly many openings converge on known openings after a few moves. I liked the no-book option of stage 1 better.

Notable game
Stockfish - Protector, round 3
Game in TCEC archive

After 23 moves the position was:



Stockfish had a small eval advantage (positional? equal material). With almost all pawns on the board not a lot could be done. The two engines took their time setting up the pawns and arranging the pieces behind them. This was the position just before the storm:



Now came a series of exchange combinations, opening files and taking the knights off the board. Both engines evaluated the position as a clear white advantage.



In the following moves Stockfish was sure the queens would be exchanged, but Protector refused to do it. Then suddenly with no warning Stockfish found the light and the eval hit the roof:



What followed was a tricky queen+pawns endgame. It took a few moves before Protector realized it was losing, and the game stopped in a 7-man position with a mate in about 20 moves according to the tablebase. As always, Stockfish was first to see it from a mile away.

Friday, September 11, 2015

Season 8 group 1b final results

The final results for group 1b are:


The six qualifiers are Stockfish, Houdini, Hannibal, Nirvana, Chiron and ... The Baron, believe it or not. Equinox and Junior were not allowed to qualify because of their multiple crashes (Pedone too but well...). This was especially sad for Equinox which played at a high level when it didn't crash. The Baron won't stand a chance in stage 2.

Crash report: Equinox crashed against Houdini late in their game, while having a small advantage as black. Junior crashed on the first move (second time this happens), forfeiting the battle for sixth place. Pedone crashed on move 9 against Junior, saving Junior from another crash.

Stockfish converted another rook ending which appeared to be drawn against Scorpio. Altogether Stockfish obtained a perfect 11/11 score in this round, remarkable. Yes, three wins were crashes but this just demonstrates the fear other engines have of Stockfish.

Crafty managed to get 1.5 points in the last two rounds, including a win against Nightmare. Not too bad for a low ranked last minute replacement.

Notable Game
Hannibal - Houdini, round 11
Game on TCEC archive

This was the battle for second place in group 1b, Hannibal in third had to win to overtake Houdini.

After 16 moves the position was:


Hannibal thought it had a small advantage. All the pawns were still on the board and white's pawn structure was unusual with two doubled pawns. Cute pattern though. A few moves and a pawn exchange later there started a long stretch of shuffling moves by the engines. Both seemed reluctant to advance pawns and the game appeared to be heading for a draw. Things started to change when Houdini took initiative and opened up the position a little.


Hannibal's forces on the king side were not enough for a significant attack, and when the dust cleared after some exchanges black had strong connected passed pawns on the queen side.



Ultimately Hannibal had to offer a bishop to get rid of these pawns, leading to a lost position.



White soon lost its passed pawns, and though it was still not clear for me how to win the two engines agreed that white was lost. Houdini managed to finish second, Hannibal a respectible third.

Thursday, September 10, 2015

Season 8 group 1b statistics

Group 1b had a 11 games that ended prematurely in an engine crash (not including the first Alfil game that was later replaced by Crafty). I did not include these games in the statistics in order to avoid biases in some of the results - draw rate is an obvious example.

For each category I briefly compare to the group 1a results. The differences represent the preferences of the engines in the two groups.

Draw rate, wins

This is quite similar to the group 1a graph, within 3% in each category.


Moves per game

median = 60.0
average = 63.67

Again similar to group 1a, medians and averages within 4 moves.


Time per game (hours)


median = 3:31
average = 3:35

Medians and averages within 4 minutes of group 1a stats.


First move


Group 1b had a significant bias for 1. e4 games, 64%. In group 1a it was 56%.

Openings

Using ECO codes there were 32 different openings. The top 3 ECO codes were:
C01 6 times - French Exchange
C12 4 times - French: MacCutcheon
C68 4 times - Spanish: Exchange

Using full opening name there were 45 different openings, a little more repetative that in group 1a. There were three openings that were repeated 3 times:
- French: Exchange, 4.Nc3 Nf6 5.Nf3
- French: MacCutcheon, Chigorin Variation
- Spanish: Exchange, 5.O-O Bg4 6.h3

Using only the first letter of the ECO codes we get the following distribution:

Compared to group 1a there were less Queen Gambit games, and more Sicillians and Indians.

If we use the opening 'family name' (using format FAMILY_NAME: VARIANT....) the top 3 are:
French - 16 times
Spanish - 8 times
Nimzo-Indian - 6 times

The French was very popular in group 1a as well, but the Spanish and Nimzo-Indian openings were very rare there.

Wednesday, September 9, 2015

Season 8 group 1b after 9 rounds

The crosstable of group 1b after 9 rounds is:


With two rounds left qualification is almost fixed. Stockfish, Houdini and Hannibal have qualified while for Nirvana and Equinox there is only a slim theoretical chance they will not qualify. The sixth qualifier will probably be Chiron. Nightmare and Scorpio have a game against Stockfish, I don't believe they have a chance. Junior has a real chance, only if it beats Chiron in the last round.

Pedone - Houdini was nearly a huge surprise. Houdini reached the endgame with an advantage of R to BP, and for many moves it seemed that both engines are shuffling pieces with no plan. According to the chat Houdini missed several opportunities to win, and made an important pawn advance only when the 50 move draw was near. In the end Houdini saw the light (perhaps Pedone blundered, it was very low on time for most of the endgame) and managed to win, but it was very close.

Hannibal - Nightmare was also a very long game. Nightmare played a rare gambit which gave Hannibal an early pawn advantage. The endgame however was very drawish and long, Hannibal made pawn moves only when close to the 50 move draw. Apparently Nightmare blundered when only opposite colored bishops remained on the board, allowing Hannibal to sacrifice the bishop for two pawns and find a win.

Nightmare played the same gambit again as black against Nirvana and lost again, perhaps not a good choice?

Crafty managed a draw against Scorpio, its first score not counting the game where Pedone crashed.

Equinox crashed twice, against Nirvana and Stockfish. It's really a shame since it is a strong engine. In the game against Stockfish the two engines were shuffling for 30 moves, and the game could have been a draw if Equinox didn't crash. Stockfish's record remains perfect.

Notable game
Houdini - Stockfish, round 8
Game on TCEC archive

This was the main attraction of group 1b, the epic battle between giants. Of course it had little effect on the qualification, and there will be many more games between these opponents in later rounds. In the context of stage 1 this was a highlight.

The game was played at a level way above my ability to describe. For the first 35 moves the evaluation scores were close to 0 for both engines. When pieces started coming off the board Stockfish started showing signs it had a small advantage, and indeed by move 49 it was up a pawn:


Was one pawn enough for this endgame?  The game progressed without changing the evaluations significantly, knights were exchanged and black added a second passed pawn on the a file.



A few moves later a pair of rooks was exchanged, as well as the passers on the d and h files, and only 7 pieces were left on the board.



Those with access to 7-man tablebases announced a draw. Houdini and Stockfish played on though, and the evaluations remained stable until around move 80 when Stockfish's evaluations started to jump. On move 84 the 7-man people announced that Houdini made a mistake and now the game was a win for black. The game continued on for another 20 moves and was adjudicated at this position:



The same 7 pieces are still on the board, but now both engines see the black win. For me it is difficult to see the difference, and the mate is still over 30 moves away. Stockfish was able to see it with 5-man tablebases, Houdini could not handle the difficult ending.

Monday, September 7, 2015

Season 8 group 1b after 6 rounds

I have only glimpsed at the games when they were live, this post is mostly from going quickly over the archives.

The situation after 6 rounds is as follows:


Stockfish, Houdini and Equinox will qualify for sure. My bet for the other three is Chiron, Hannibal and Nirvana. Nightmare is also a surprise contender with low rating. I don't think any of the others has any chance, including Junior which is having a bad tournament with too many crashes.

Junior recorded the quickest crash possible when it crashed as white before making a move against Stockfish. Some say it was too scared to play. And this was with reduced threads and hash to "improve stability".

Stockfish had two impressive wins against strong opponents, Hannibal and Nirvana. So far I am happy with its performance, no time trouble and a good attacking style. The real tests are still to come in the following stages.

In rounds 5 and 6 Crafty's evaluation was somehow inverted. It played as black and gave positive scores for black's advantage, contrary to the convention of TCEC. As a result the GUI could not adjudicate when white was winning since Crafty gave high negative scores. The games played on until a TB adjudication was possible. This was later fixed.

Notable game
Nightmare - The Baron, round 4
Game on TCEC archive

The game was balanced from the start, with evaluation scores close to 0 from both engines. We join the action after move 24.



The Baron saw an opportunity to take the f2 pawn, but this allowed Nightmare to trap the black rook. After exchanging pawns on the queen side Nightmare created a passed pawn on the c file, backed by a rook.



The a pawn was next to go, Nightmare had two passed pawns backed by rooks. The trapped rook broke free but only to aid in the blocking of the white pawns.



White now used its queen to break the blockade preventing the white pawns advance to the 8th row.


What a beautiful position! White threatens to take the rook, and black cannot take the queen because it loses the rook.The game was over within a few moves.

Saturday, September 5, 2015

Season 8 group 1b after 3 rounds

It's been a busy week and I didn't follow all the action in the 1b group very closely, these are my impressions so far.

Group 1b standings after 3 rounds:

Unfortunately the issue that was in focus in the first rounds was the crashes. There were too many of these, especially in contrast to the 1a group.

In the Pedone - Alfil, Alfil replied immediately from book, and Pedone crashed on move 2. Since books are not allowed Alfil was disqualified and Crafty replaced it - that's why you don't see Alfil in the table. However, when the game was replayed with Crafty, Pedone crashed on move 7. Pedone's troubles continued when it crashed against Stockfish in round 2, on move 9 this time. After limiting the number of threads and hash size it managed to complete a game in round 3, hopefully stable enough to play now.

Junior is also causing problems with 2 crashes against Houdini and Nirvana in the first three rounds. Martin wrote in the chat that he was certain that this version of Junior wasn't stable enough and recommended using a prior version, but he was overruled (developers? viewers? someone insisted on the new version).

Stockfish with 3 wins out of 3 (one was a crash), it seems to be back after a problematic season 7. Still early though, and the opponents are too weak for a real assessment.

Chiron - Nirvana was a miniature, ending in 3-fold repetition after 11 moves.

Notable game
Hannibal - Junior, round 1
Game on TCEC archive

Junior's 6th move allowed Hannibal to advance to e5 attacking a knight:


This led to an exchange of knights and the opening of the king side:



Black is a pawn up temporarily, but has no development on the queen side and its king pawns are undefended. Junior tried to attack the white king, forcing it to advance. The king soon found a safe spot, allowing Hannibal to eliminate all of black's king side pawns.



White now was a pawn up and had two passed pawns, black still had development problems on the queen's side. The queens were exchanged next, and white gained another pawn. The rest of the game was a straightforward case of exchanging pieces until the advantage is converted to a win. Well done Hannibal !

Wednesday, September 2, 2015

Season 8 group 1a statistics



Draw rate, wins


Moves per game

median = 56.75
average = 65.93


Time per game

median = 3:35
average = 3:39

First move


Openings
Using ECO codes there were 39 different openings. The top 3 ECO codes were:
C01 7 times - French Exchange
D37 6 times - QGD
C54 5 times - Giuoco Pianissimo

Using full opening name there were 58 different openings, so almost no repeated opening. The QGD: Classical, 5...O-O was the only opening repeated 3 times

Using only the first letter of the ECO codes we get the following distribution:

Very few Indians and Sicillians.

If we use the opening 'family name' (using format FAMILY_NAME: VARIANT....) the top 3 are:
French - 18 times
QGD - 12 times
Two Knights - 4 times