Stockfish wins the season 19 superfinal. The final result is 18-9 with 73 draws. In the last 10 games Stockfish increased its lead by two game pairs.
The superfinal started with relatively a few decisive game pairs, after 50 games Stockfish led 5-4 and only 3 game pairs were decisive. In the second half of the match the draw rate dropped to 64% and the score was 13-5. I think Jeroen mentioned in the chat that the opening bias was designed to increase as the match progressed. Still it was unexpected that Leela did not win any game pair in the second half while Stockfish won 8.
This season Stockfish used NNUE, with a neural net for evaluation (on CPU I think? not sure). The NN revolution that started with AlphaZero, and Leela's first participation in TCEC season 12 two and a half years ago, appears to be complete. This season both TCEC superfinalists, indeed the top 4 engines in the premier league, all used neural nets for evaluation. Other engines switched to NNUE and there may be more new NN engines being developed. How long until there are no "classical" evaluation engines playing in the premier league?
In the lower leagues we saw the rise of SlowChess, a new non-NN engine that started from qualification and reached league 1. Igel started in league 3, then transformed to Igel NN and this boosted it to league 1 as well. In the premier league Stockfish and Leela led from early on and most of the attention was the race at the bottom of the league. Four engines battled against relegation, including former champion Komodo. In the last DRR Komodo moved up with a few rare wins, Ethereal dropped to last while Fire and ScorpioNN were tied in the last round. In the end it was Fire that relegated and ScorpioNN survived.
I suppose the added NN improved Stockfish's opening abilities, this was always the strong point for Leela. Looking at some of the games of the superfinal suggests that Stockfish NNUE and Leela made similar choices in the opening, much more than before. In my statistics report you can see that the distribution of equal moves after book is much more flat than ever before in TCEC, so the game pairs repeated more moves than usual. In games 21-22 with a 2-ply book the games diverged after 21 plys. Games 25-26 also had a 2-ply book, the games diverged after 25 plys ignoring an early move transposition. The engines are still not the same, in games 45-46 there was a 2-ply game and the games diverged after only 6 plys.
On a personal note, in season 19 the blog reached 100K all time views. I enjoy looking at the geographic distribution of readers, spread all over the world. This is not something I imagined when I started it 5 years ago. Thank you for reading, I welcome comments and corrections.
The corona virus is still with us, this week there was a new record of ~500K people testing positive in a day. Please take this seriously, avoid crowded places and unnecessary contacts. Keep yourselves and your family safe. Governments may give bad advice or act slowly, but our personal health depends mostly on our personal behavior.
See you next season,
In game 91 Leela developed its rook through the h file and didn't castle, while Stockfish castled long. There were only a few exchanges after the start and the engines locked pawns in the center and queen side. On move 30 Leela opened a file on the king side and evals increased over 1.5. The engines exchanged pieces through the gap until the game reached a RN vs RB position. Leela's eval was over 2 but it couldn't break through the black pawn wall. The engines shuffled and Leela moved its f pawn reluctantly while the evals came slowly down. On move 103 Stockfish took the f pawn and evals dropped close to 0, after a few pawn exchanges the game was adjudicated. Did Leela miss a win?
In game 92 Stockfish castled and Leela kept its king in the center. Leela pushed a pawn on the queen side, a pawn exchange opened a way for a white rook to move forward. Stockfish thought that Leela's move 21 was a mistake, it should have tried to castle its king. When it didn't Stockfish's eval jumped, it sacrificed a knight and forced the black king to move.
Leela's eval reacted 5 moves too late. Stockfish doubled rooks and attacked the a pawn, after a series of exchanges the game reduced to a Q vs QB position and Stockfish had 5 extra pawns for the piece.
Leela had too many pawns to worry about, and its king had no protection except the bishop. Stockfish let the 2 pawns on the queen side be captured while it moved its king forward, finding protection from the black queen that was behind the pawn line.The e pawn marched forward and Leela could not stop it, it lost its queen and the game. Stockfish wins the game pair and its lead increases to +8.
In game 93 there were no exchanges after the start, Stockfish formed a pawn line across the board. After most of the minor pieces were exchanged the engines shuffled for a while, then Leela opened a file on the queen side. The engines exchanged all rooks through the gap in the wall and the game reached a QB vs QB position on move 28. Stockfish gave a pawn and moved its queen forward where it could give checks. Evals were close to 0 and the draw rule stopped the game quickly.
In game 94 Leela formed a pawn line as well but it was less stable than in the previous game. Stockfish exchanged a pair of pawns on move 16 and opened a gap in the line, after a series of exchanges the position opened. Stockfish was a pawn up with a central passer though its pawn structure looked awful. Both kings were vulnerable on the open g file. Stockfish thought it was better with an eval over 2.
Leela placed a rook on the g file first. Stockfish exchanged pawns and removed the black pawn on f5, making its f pawn a passer. Leela blocked the second passer with its rook and Stockfish took over the g file. By the time Leela captured the f pawn Stockfish lined up all 3 major pieces on the g file.
Leela needed all available pieces to defend its king, and one rook was too far to help. Stockfish moved its queen forward and pushed the passer, Leela could not stop it. Stockfish got a bishop for the passer, mate was coming when the game was adjudicated. Stockfish wins another game pair, increasing its lead to +9.
Games 95-96 started with a 20-ply book in the Vienna opening, where the black king moved without castling and white captured a rook with a knight, which is trapped in the corner. In game 95 Stockfish captured the knight and Leela castled its king. Stockfish pushed 4 pawns on the king side, it gave one away but managed to trap and capture a knight. Stockfish's eval was 0, it opened the king side and attacked the white king. After a series of exchanges only RR vs RBN remained on move 38. The engines continued to exchange pieces and pawns, the game ended in a tablebase draw. Game 96 repeated the same moves for 30 plys, except for a few early transpositions, resulting in a R vs NN imbalance and advanced black pawns on the king side. Leela opened the king side and evals were close to 0. The engines reduced to a rook ending with Stockfish a pawn up, the pawns did not move and the game was quickly adjudicated.
Games 97-98 started with a 26-ply book, a high level line in the King's Indian defense, orthodox variation. The center was locked and there were no early exchanges. In game 97 the engines exchanged a pair of pawns and a pair of knights on the queen side, but made no attempt to attack. Stockfish opened a file on the king side and the engines exchanged a pair of rooks. Evals were close to 0, the pawns did not move and the draw rule ended the game quickly. In game 98 Stockfish attacked on the queen side, it captured a pawn and its eval jumped over 1.5. Leela pushed pawns on the king side and gave another pawn, Stockfish's eval dropped back to 0. Leela sacrificed a knight and opened the king side, a queen and a knight were sufficient to force a perpetual check draw.
In game 99 there were a few minor piece exchanges after the start, Stockfish went a pawn up and its eval dropped to 0. The engines started to shuffle on move 25, Stockfish was content in defending its fortress and Leela made no attempt to attack it. There was one pawn move until the game was adjudicated on move 92. In game 100 Stockfish castled long, Leela gave a pawn and tried to attack on the queen side. Stockfish was not worried, its eval dropped to 0 as it exchanged a pair of rooks and created a central passer. Leela was busy blocking the passer, its eval was low enough for the game to end quickly by the draw rule.