Monday, June 22, 2020

Season 18 superfinal games 1-10

After 10 games Stockfish leads 4-1 with 5 draws. This result is quite unexpected, considering that Leela beat Stockfish last season, not too long ago. The chat is buzzing, with fans of the two engines arguing and trying to figure out what is going on. Various explanations have been suggested, including the engine software updates, Leela's configuration and the upgraded hardware. It could be a random result of the small sample size, likewise for the result of the previous superfinal, but who wants to believe in statistics? Another contributing factor is the openings. As Jeroen explained, the openings for this superfinal are more biased as he preferred more decisive games, at the cost of having more openings with wins for both engines. Perhaps Stockfish is better in such openings? More decisive games means more work for me, but I'm not complaining too hard.

Games 1-2 started with a Sicilian 2-ply opening. Game 1 continued in the Kan Polugaevsky variation. There were very few exchanges and evals stayed low. Stockfish started shuffling before move 20, after a few exchanges on move 29 both engines continued to shuffle and the game was adjudicated. In game 2 the engines played a B50 Sicilian with d6 and d3 played. There were many early exchanges and only QRB vs QRN remained on move 24. Evals were close to 0 and the game was adjudicated before move 40. 

Leela had an eval advantage from the start of game 3, Stockfish seemed to be holding in the first 20 moves. The center was blocked, Leela pushed pawns across the board and kept its king uncastled. The first exchange was on move 20, Stockfish prepared a king side attack and Leela concentrated on the queen side. Stockfish's eval went over 1, on move 27 it moved its queen away from defending the b7 pawn, but it didn't expect Leela to take it. 

Leela captured the pawn and Stockfish's eval immediately jumped as it saw the blunder (so why did it make the move in the first place??). Leela traded Q for RR, then a quick series of exchanges led to a RB vs BN ending, white a pawn up.

Stockfish had a passer that had to be watched but the material was enough for Leela to win.

In game 4 Stockfish pushed pawns on the king side and castled long. It prepared a king side attack, doubling rooks on the g file, and its eval went over 2. Leela refused a bishop sacrifice but it couldn't prevent the king side being opened. Leela moved its pieces to surround the king, that was left without pawns.

Stockfish shifted its rooks to the open e file, and created a passer on the queen side to increase the pressure. It gave a rook for a knight and prepared for the final attack. 

The white rook and queen moved forward, Leela defended its king and exchanged pieces until only QN vs QB were left. Stockfish was only a pawn up, but after Leela ran out of checks Stockfish forced a queen exchange and the b passer was unstoppable. There were two white wins in this opening, score is 1-1.

Stockfish moved its king early in game 5, out of known theory. Evals increased over 1, the engines opened a file on the queen side, Stockfish stabilized its pawns and started to shuffle. Leela couldn't find an entry point, it offered a knight but Stockfish wouldn't open the king side. Evals slowly came down, after exchanging a few pieces Leela started to shuffle as well, the game ended with a 50-move draw.

Game 6 continued along a more popular line, Stockfish's eval was around 1.5. Leela gave a pawn early to activate its pieces and create a king side threat. Stockfish concentrated on the queen side, it opened a file and placed a knight on c5. Leela created a queen side passer and gave a rook for a knight and pawn to remove the white knight. Stockfish's eval was over 2.

Stockfish moved its heavy pieces to the queen side, it blocked the passer with a rook while taking over the open file. Leela played passively and evals continued to increase .On move 41 Stockfish exposed the black king with a rook sacrifice. 

The white queen came forward, it chased the black king while capturing pawns. Stockfish was 3 pawns up when it exchanged queens with only RB vs RNN remaining. Leela couldn't stop the white king side pawns marching and the game was adjudicated. Stockfish wins first game pair and goes into the lead. 

There was a series of early exchanges in game 7, Leela was up a rook for a knight while Stockfish was two pawns up with connected passers on the queen side. Stockfish's eval came down to 0, Leela's king blocked the passers while its pieces couldn't find a way through the black defense. The position opened with more exchanges, Stockfish gave a piece and its queen side passers and Leela was left without pawns in a RB vs N position. Stockfish had two connected pawns, close enough to promotion so they were worth a rook. The game ended in a tablebase draw. 

In game 8 there were more pieces on the board. Stockfish thought Leela's 15th move was a mistake and its eval jumped over 1.5. It immediately gave two minors for 3 pawns and its eval jumped higher while Leela was completely calm. After exchanging queens the center opened, Stockfish captured two more pawns and created a queen side passer. The white rooks controlled the center and the black king was exposed, yet Leela's eval remained close to 0.

Leela realized something was wrong only on move 27. It gave back some of the material but couldn't save its king. Stockfish attacked with its rooks and managed to capture a black rook.

Leela's last hope was its passer, Stockfish traded a rook for a bishop to get rid of it. The rook ending with 3 pawns up was as easy win. Such engine blindness in a superfinal is a surprise to me, I don't think there was such an extreme case in the last several seasons. Stockfish wins the game pair and increases the lead to +2.

Games 9-10 started with a King's Gambit accepted, with black up a knight for a pawn. In game 9 there were a few early minor piece exchanges. Stockfish was late in developing its king side and its king was uncastled in the center. Leela moved its heavy pieces to the half open f file, Stockfish prepared an escape route for its king to the queen side. Stockfish's (negative) eval increased steadily and was over 2.5 on move 26.

Leela moved its rooks forward, Stockfish captured pawns while Leela shifted its pieces to threaten the king on the queen side. In a series of exchanges Leela regained the piece and reduced to a queen ending with black a pawn up. Despite the small material advantage Stockfish's eval was close to 20 by then, Leela was more hopeful.

The ending was far from trivial and the game continued for more than 25 moves. After a few pawn exchanges Stockfish pushed its queen side pawns forward. The black king had more cover and the white king was exposed to checks, Stockfish used this to get a pawn to the 2nd rank supported by its king. When the second pawn moved forward the game was adjudicated.

In game 10 both engines castled long early and evals came down. The black queen moved forward, Leela traded it for a rook and bishop to get a Q vs RBN imbalance on move 17. The white queen was very active, Leela's pieces defended each other and did not try to attack. Stockfish opened the queen side and threatened the black king, the engines mostly shuffled and the game was adjudicated when the pawns stopped moving. Stockfish wins another game pair and the lead is +3.

No comments:

Post a Comment