TCEC Season 15 Fun Bonus Game 2 – Leela-Stockfish

From TCEC wiki
Revision as of 15:09, 13 October 2019 by Skiminki (talk | contribs)
(diff) ← Older revision | Latest revision (diff) | Newer revision → (diff)
Jump to: navigation, search
Game Season 15, Fun Bonus, Game 2
Result 1-0
ECO E06
Links TCEC archive
Lichess



This game was played during TCEC Season 15 in a bonus event. The bonus event consisted of 8 games between Leela and "Bluefish." Bluefish was a kibitzer engine at the time running Stockfish on strong hardware (4x Intel Xeon E5-4669v4[1]), totaling 88 cores and 176 threads. Leela was running on a Jhorthos net T40.T6.532[2]. This Jhorthos net was a predecessor to the one that played in the Superfinal of Season 15.

This game stands as a showcase of the positional strength of Leela. Leela, playing as white, takes advantage of a single inaccurate move by Stockfish, essentially gifting a single critical tempo to Leela. Leela uses this opportunity to impose relentless pressure and accumulating small positional improvements with the help of the pressure. This is done by shifting the focus of the game from one side of the board to another, and to force piece trades with favorable conditions. Black always seems to be one or two tempi short on stabilizing but is never given the opportunity. Finally, white has accumulated enough small gains for a winning position.

Game overview

    
    
    
    
    
    
    
    
    
   
      
     
     
       
  
     
 ♞♝  ♜♚ 
♜   ♝♟♟♟
♟   ♟   
♘ ♟    ♛
♙♟♘     
      ♙ 
 ♙♕ ♙♙♗♙
  ♖♖  ♔ 
Position after 18. Nba5


This game was played from the starting position.

1. d4 Nf6 2. c4 e6 3. g3 d5 4. Bg2 Be7 5. Nf3 O-O 6. O-O dxc4 7. Qc2 b5

Instead of following the mainline theory with 7... a6 offering back the pawn, black opted to hang on to being pawn up. However, as a compensation, white gets some activity and easier development.

8. a4 b4 9. Nfd2

9. Nfd2 may be a bit odd-looking move at first, but it is well-trodden paths in opening theory. The merits of the move include activating the bishop on g2 and undermining the b4/c4 pawns. Black is now essentially forced to play 9... c6, which hinders the queenside development.

9... c6 10. Nxc4 Qxd4 11. Rd1 Qc5 12. Be3 Qh5 13. Nbd2 Nd5

This is where things started to go a bit wrong for black, and 13... Ng4 was the primary move to consider. While both knight moves have the same goal to take the bishop on e3, the additional Qxh2+ threat associated with 13... Ng4 makes all the difference. To parry the threat, white would have had to bring the knight on d2 to the kingside by 14. Nf3 or 14. Nf1, or weaken the kingside with 14. h4 or 14. h3. But in the game continuation, white now gets a free tempo, which allows white to strengthen the grip on the queenside with 14. Nb3. At this level, every tempo counts.

As one could have expected, 13... Nd5 was also a novelty in the Lichess Masters database.[3]

14. Nb3 Nxe3 15. Nxe3 a6 16. Nc4 Ra7 17. Rac1 c5 18. Nba5 (diagram)

Black's problems in the queenside development are now exemplified: the knight cannot move and the light square bishop cannot challenge the a8-h1 diagonal. An attempt to develop the bishop with 18... Bd7 would further disrupt the coordination with the black rooks. Black cannot really challenge the d-file either with 18... Rd7 or 18... Rd8, since white can simply take and then black would have a clumsy minor piece the d-file. White's two knights and the bishop on g2 properly dominate black's queenside, giving white a significant advantage.

White now changes the focus on the kingside, taking advantage of black's problems in queenside piece development.


    
    
    
    
    
    
    
    
      
    
    
      
     
    
       
       
    ♜ ♚ 
   ♖♝♟ ♟
♟♘  ♟  ♙
  ♟    ♞
♙♟   ♙  
 ♙  ♙♙♟ 
      ♗ 
      ♔ 
Position after 38... Re8


18... Qg5 19. h4 Qf6 20. Qe4 Rc7 21. Rd3 g5 22. h5 Qg7 23. h6

The thorn pawn on h6 is now installed. It cannot be taken, because black would lose a piece on the skewer: 23... Qxh6 24. Qe5 Rd7 25. Qxb8.

23... Qf6 24. b3 Bd7 25. Qe3 g4 26. Rcd1 Be8 27. Kh2 Qg5

Black understandably wants to exchange the queens to relieve some of the attack pressure. White obliges, but dictates better terms with 28. Qf4, skewering the rook on c7 and the knight on b8. Black's best option is to trade the queens on white's terms.

28. Qf4 Qxf4 29. gxf4

After the queen trade on f4, white gets a powerful pawn on f4 controlling important squares on e5 and g5. This control prevents black from targeting the annoying thorn pawn on h6 any time soon.

29... Nd7 30. Nc6 Bh4 31. Kg1 Nf6 32. N6e5 Nh5 33. e3 g3 34. f3

After the kingside is now in white's control, the attention is shifted back to the queen side.

34... Be7 35. Nd7 Bxd7 36. Rxd7 Rfc8 37. Nb6 Rxd7 38. Rxd7 Re8 (diagram)

Now white has multiple ideas to dismantle black:

  1. Take black's a-pawn to liberate a queenside passer.
  2. Plant bishop on c4 to start pressurizing f7 and e6. Pawn to f5 can be played to strengthen this idea.
  3. Knight can be rerouted to a better square to further pressurize f7, for instance.
  4. If black's dark square bishop gets deflected or exchanged, then black's c5 and b4 pawns become weak.
  5. The pawn on g3 is ripe for picking.

Black's ideas for counterplay are much more limited because the pieces are tied in protecting the pawns. However, white's h6 pawn is currently weak and some effort must be paid to take it now when there is the chance.

None of white's ideas alone are enough to win the game. But combined, they are objectively winning the game for white.


    
    
    
    
    
    
    
    
      
      
      
     
    
     
        
        
  ♜  ♚  
     ♟ ♟
   ♝♟   
♖ ♟ ♙   
♙♟♗  ♙  
 ♙   ♙♔ 
        
        
Position after 51. e5


39. Bf1 Nf6 40. Rb7 Nd5 41. Nc4 Bf8 42. Kg2 Bxh6 43. Kxg3 Idea 5: The first piece of the puzzle is now completed.

43...Bf8 44. e4 Ne7 45. Nd6 Idea 3: Knight improved. Now threatening to take the pawn on a6.

45... Ra8 46. Rd7 a5 47. Bc4 Idea 2: Bishop on c4. The threat is now pawn to f5 winning a crucial black pawn to create a passer on the e-file or f-file, so black offers a knight trade.

47... Nc8 48. Nxc8 Rxc8 49. Ra7 Idea 1: After the position has simplified a bit, black cannot prevent white to take the a-pawn anymore.

49... Bd6 50. Rxa5 Kf8 51. e5 (diagram) Idea 4: Deflection of black's bishop. Now white forces two passed pawns whit couple of files in between. In general, two passed pawns with some horizontal distance is enough to win an opposite colored bishop ending, as long as neither of the pawns is on the border row with the wrong color promotion square. The distance needed depends on which rows and how advanced the pawns are, but generally 3 empty files between the pawns is enough, and less, when the pawns are advanced.

The black bishop now has three squares to go, but none of them work:

  • 51... Be7 52. Ra7 Ke8 53. a5 Kd8 54. f5 exf5 55. Kf4 Rc7 56. Ra6 h5 57. Rh6 Rd7 58. Rxh5 and white will get passed pawns on both sides of the board. In this case, the opposite colored bishops cannot secure a draw.
  • 51... Bb8 52. Rb5 Ke8 53. a5 Kd8 54. a6 Ba7 55. f5 exf5 56. Rb7 Rc7 57. Rxc7 Kxc7 58. Bxf7 and white again has two passed pawns with enough distance to avoid opposite colored bishop draw.

So, black decided to play Bc7 and give up the pawn on c5 voluntarily. But there is no stopping of white's pawn grabbing and pawn advancement anymore.


    
    
    
    
    
    
    
    
        
      
     
        
        
       
        
        
        
  ♚♗    
    ♙♙♔ 
        
        
♝       
        
        
Position after 100... Kxc7


51... Bc7 52. Rxc5 Ke7 53. a5 Kd8 54. a6 Bb6 55. Rb5 Ba7 56. f5

Now white creates the second passed pawn for white. The rest is simple technique. However, one thing to note is that in this game, Leela was configured with only the 6-men Syzygy[4] WDL tables, but not with the DTZ ones. This was because usually the games are automatically adjudicated at 6-men positions and the DTZ tables are not needed.

56... Rb8 57. Ra5 exf5 58. Bxf7 Bb6 59. Ra2 Rc8 60. Kf4 h5 61. Kxf5 Rc7 62. Bxh5 Rh7 63. Bg4 Rc7 64. Ra4 Ra7 65. Rxb4 Rxa6

Here is a forced mate in 21: 66. Ra4 Ra5 67. Rxa5 Bxa5 68. Ke6 Ke8 69. f4 Kf8 70. f5 Kg7 71. f6+ Kg6 72. Kd7 Bb4 73. Ke8 Kg5 74. f7 Kxg4 75. e6 Kf5 76. e7 Ke6 77. f8=Q Ke5 78. Kf7 Bxe7 79. Qxe7+ Kd4 80. Qe2 Kc3 81. Ke6 Kxb3 82. Kd6 Kc3 83. Kc5 Kb3 84. Qd2 Ka3 85. Kc4 Ka4 86. Qb4#

66. f4 Bc5 67. Rc4 Bf2 68. Ke4 Rg6 69. Kf3 Bh4 70. b4 Ra6 71. Bf5 Be1 72. b5 Ra3+ 73. Kg4 Rg3+ 74. Kh5 Ba5 75. Ra4 Bb6 76. Bg4 Ke8 77. Kg5 Bd8+ 78. Kf5 Bb6 79. Ra6 Bf2 80. Bh5+ Ke7 81. Re6+ Kd7 82. Be8+ Kc7 83. Rc6+ Kb8 84. Rc4 Ka7 85. e6 Rg7 86. Ke5 Bh4 87. f5 Be7 88. Bg6 Bf8 89. Rc6 Rb7 90. Kf6 Be7+ 91. Kg7 Ba3+ 92. Kh6 Rxb5 93. f6 Bf8+ 94. Kh7 Rb7+ 95. Bf7 Ba3 96. Rc2 Kb6 97. Kg6 Kb5 98. Be8+ Kb6 99. Bd7 Rc7 100. Rxc7 Kxc7 (diagram)

With the help of the WDL tables, Leela found relatively quickly to a winning 6-men database position. However, without the DTZ tables, Leela is on her own. In completely winning positions, Leela does not care about how quickly the game ends as long as it ends in a win. With perfect play on both sides, the game would have ended in about 20 moves.

101. Kf7 Bc5 102. e7 Kxd7 103. e8=Q+ Kd6 104. Kg8 Kd5 105. Qa8+ Kd4 106. Qc8 Bb4 107. Qd7+ Kc3 108. Qe8 Kd4 109. Qf7 Kc3 110. Qg7 Bc5 111. Qh8 Kc4 112. Qh7 Kc3 113. Qg7 Kc4 114. Qh8 Bd4 115. Qh7 Bc5 116. Qg6 Kc3 117. Qf7 Bd6 118. Qe8 Bc5 119. Kg7 Kd3 120. Kg6 Kd4 121. Kf5 Bb4 122. f7 Kc4 123. Qg8 Kb5 124. Qc8 Bc5 125. Qb8+ Kc4 126. Qa8 Bb4 127. Qa7 Kd3 128. Qa6+ Kd4 129. Qa8 Kc4 130. f8=N Kd3 131. Ng6 Kd4 132. Nf4 Be1 133. Qa7+ Kc4 134. Qb7 Bc3 135. Qc8+ Kb3 136. Qd7 Bh8 137. Qe8 Bc3 138. Qf7+ Kb4 139. Qg8 Be1 140. Qh7 Bd2 141. Qg6 Kc5 142. Qe8 Kd4 143. Qf7 Kc3 144. Qg8 Be1 145. Qh7 Bd2 146. Qg6 Kd4 147. Qe8 Kc5 148. Qa8 Be1 149. Qa7+ Kb5 150. Qb7+ Kc4 151. Qc7+ Kb5 152. Qd7+ Kb4 153. Qe7+ Kb3 154. Qxe1 Kc4 155. Qe8 Kc5 156. Qd8 Kc4 157. Qf8 Kb5 158. Qc8 Kb6 159. Qd8+ Kc5 160. Qe8 Kc4 161. Qf7+ Kb5 162. Qg7 Kc6 163. Qh7 Kb5 164. Qh6 Kc5 165. Qh5 Kd4 166. Qh4 Kc3 167. Qh3+ Kc2 168. Qh8 Kb3 169. Qb8+ Kc4 170. Qa8 Kc3 171. Qb7 Kd4 172. Qa7+ Kc3 173. Qa6 Kb4 174. Qb6+ Kc4 175. Qc7+ Kb4 176. Qd7 Kc5 177. Qe7+ Kc4 178. Qd8 Kc3 179. Qd7 Kc4 180. Qe7 Kc3 181. Ke5 Kc4 182. Kd6 Kd4 183. Qd8 Ke4 184. Qf8 Kd4 185. Qc8 Ke4 186. Qd7 Kxf4 187. Qc8 Ke4 188. Qd8 Kf4 189. Qb8 Ke4 190. Qa7 Kf4 191. Qa6 Ke4 192. Qa5 Kf4 193. Qa4+ Ke3 194. Qa8 Kf4 195. Kd5 Kf5 196. Qa7 Kf6 197. Qb7 Kf5 198. Qb8 Kf6 199. Kd6 Kf5 200. Qa8 Kf4 201. Ke6 Ke3 202. Ke5 Kd3 203. Qa7 Kc3 204. Qb8 Kc4 205. Kd6 Kd4 206. Qe8 Kc3 207. Kc5 Kc2 208. Kc4 Kd2 209. Qe7 Kc1 210. Kc3 Kd1 211. Qe8 Kc1 212. Qe1# 1-0

Editor's note

    
    
    
    
    
    
    
    
     
  
      
     
    
      
     
      
 ♞   ♜♚ 
  ♝♝♛♟♟♟
♟   ♟   
♘ ♟ ♙   
♙♟♘    ♙
 ♙    ♙ 
  ♕  ♙♗ 
  ♖   ♔ 
Position after 23... Bc7
Variation: 18... Bd8


It is hard to pinpoint the exact position of no return for black. The game diverges from Catobase on move 18... Qg5. As per Catobase, 18... Rc7 has been tried (+0 =0 -1) as well as 18... Bd8 (+0 =2 -0). The point of divergence indeed seems to be a critical position.

After 18. Nba5, the computer analysis suggests the following evaluations for black's top choices by SF: (SF 68.5B nodes, depth 42; Leela 1M nodes per move)

  1. 18... Bd8: SF +0.29; Lc0 64.1%
  2. 18... Bg5: SF +0.47; Lc0 65.3%
  3. 18... Rc7: SF +0.50; Lc0 67.4%
  4. 18... g6: SF +0.57; Lc0 68.5%

Additionally, Leela suggests:

  • 18... Rd7: SF +0.57; Lc0 65.2% (SF 4.9B nodes, depth 38; Lc0 1M nodes)

Playing couple moves based on their best moves (more weight for Leela's suggestions for white, and similarly more weight on SF for black) the following continuation is suggested: 18... Bd8 19. b3 Qg5 20. e4 Rd7 21. Rxd7 Bxd7 22. h4 Qe7 23. e5 Bc7 (diagram). Here, the Leela evaluation is 63% (1M nodes) and Stockfish evaluation is +0.86 (SF 12B nodes, depth 41). White still has the edge but perhaps black is able to stabilize.

The move played in the game was evaluated as:

  • 18... Qg5: SF +0.67 (18.7B nodes, depth 38); Lc0 (67.7% 2M nodes)

Engines used: SF-dev (8fec8834), multi-PV 4; Leela-dev (db42c60) with T40B.4-160.

Analysis

Kingscrusher's game analysis