AlphaZero SF game 10 Does Your Engine Find 29Qh3?

[carldaman]

By the way, Ed, I really liked your approach with Prodeo 2.2, using different cores to run separate searches. I was hoping more programmers would follow suit and improve on that approach, but it didn't quite happen. There seems to be an entrenched resistance to innovation in the whole chess programming field and only the AZ surprise may have awakened some folks.

The idea using a second (somewhat speculative) search was not meant to gain playing strength but to improve on playing style (hopefully) without ELO loss. That it eventually turned out a small improvement was an unexpected case of collateral damage in reverse.

I think that most programmers would call it a waste of time because a typical doubling should bring 30-35 ELO and not only 10 ELO.

They're too Elo conscious, Ed, and have forgotten what style is -- why we play this game in the first place!

[Eelco de Groot]

Hi Eelco
Just curious,
[D] rn2r1q1/p5k1/2p3p1/1p4p1/2PR4/6PQ/5PK1/7R b - -
after 34 ... bc4
what does kaissa think of the black position after 35 g4? (instead of (Qh6+)
Best Fonzy

At large depths, one move earlier in the game Kaissa now can predict 9 consecutive moves of Aleph-Null correctly, then choosing another move than SF did so the string ends. The move from Aleph is once again a singular move, although in a previous search I got a big score for 33. Rd6 but not anymore now, after Black chooses another move 33... bxc4 as reply. I assume 33. Bd4 is actually singular but the 'proof' is circumstantial. These positions are very difficult for Stock to judge. So I think this version might be ever so slightly better than the other one.

[D]rn2r1q1/p5k1/2p2bp1/1p4p1/2P5/4B1PQ/5PK1/3R3R w - -

Engine: Kaissa search_010, bench 5724923 (one thread, 512 MB)
by T. Romstad, M. Costalba, J. Kiiski, G. Linscott

39 164:51 +1.59 33.Bd4 Bxd4 34.Rxd4 Rd8 35.Rxd8 Qxd8
36.Qe6 Nd7 37.Rd1 Nc5 38.Rxd8 Nxe6
39.Rxa8 Kf6 40.cxb5 cxb5 41.Kf3 {This is all as in the game} Ke5
42.Ke3 Ng7 43.Rxa7 Nf5+ 44.Kd3 Kd5
45.Rb7 Kc5 46.Rb8 (24.737.369.501) 2500

39 164:51 0.00 33.Qh6+ Kf7 34.Bxg5 Qg7 35.Qh4 Re5
36.Bxf6 Qxf6 37.Qh7+ Qg7 38.Qh4 Qf6 (24.737.369.501) 2500

39 164:51 0.00 33.Rd6 bxc4 34.Qh6+ Kf7 35.Rxf6+ Kxf6
36.Qxg5+ Ke6 37.Rd1 Qf7 38.Qg4+ Qf5
39.Qxc4+ Kf6 40.Bd4+ Ke7 41.Bc5+ Kf6
42.Bd4+ (24.737.369.501) 2500

39 164:51 0.00 33.Rd2 a5 34.Rd6 Be5 35.Rdd1 a4
36.Bxg5 Kf7 37.Rd3 Qg7 38.Qg4 Kg8
39.Bh6 Qf7 40.Rf3 Bf6 41.Bg5 Nd7
42.Qh4 Qg7 43.Rxf6 Nxf6 44.Bxf6 g5
45.Qd4 Qg6 46.cxb5 (24.737.369.501) 2500

39 164:51 -0.08 33.Rd3 bxc4 34.Rd6 Be5 35.Rdd1 c5
36.Qh6+ Kf7 37.Qxg5 Re6 38.Bxc5 Bf6
39.Qd5 Nc6 40.Bd6 Qc8 41.Rh7+ Bg7
42.Re1 Ne7 43.Rxg7+ Kxg7 44.Qxe6 Qxe6
45.Rxe6 Nd5 46.Re5 (24.737.369.501) 2500

38 164:51 0.00 33.cxb5 Re5 (24.737.369.501) 2500

[pgn]


[Event "?"]
[Site "?"]
[Date "????.??.??"]
[Round "?"]
[White "AlphaZero"]
[Black "Stockfish"]
[Result "1-0"]
[ECO "E17"]

1.Nf3 Nf6 2.d4 e6 3.c4 b6 4.g3 Bb7 5.Bg2 Be7 6.O-O O-O 7.d5 exd5 8.Nh4 c6

9.cxd5 Nxd5 10.Nf5 Nc7 11.e4 d5 12.exd5 Nxd5 13.Nc3 Nxc3 14.Qg4 g6 15.Nh6+

Kg7 16.bxc3 Bc8 17.Qf4 Qd6 18.Qa4 g5 19.Re1 Kxh6 20.h4 f6 21.Be3 Bf5 22.

Rad1 Qa3 23.Qc4 b5 24.hxg5+ fxg5 25.Qh4+ Kg6 26.Qh1 Kg7 27.Be4 Bg6 28.Bxg6

hxg6 29.Qh3 Bf6 30.Kg2 Qxa2 31.Rh1 Qg8 32.c4 Re8 33.Bd4 Bxd4 34.Rxd4 Rd8

35.Rxd8 Qxd8 36.Qe6 Nd7 37.Rd1 Nc5 38.Rxd8 Nxe6 39.Rxa8 Kf6 40.cxb5 cxb5

41.Kf3 Nd4+ 42.Ke4 Nc6 43.Rc8 Ne7 44.Rb8 Nf5 45.g4 Nh6 46.f3 Nf7 47.Ra8

Nd6+ 48.Kd5 Nc4 49.Rxa7 Ne3+ 50.Ke4 Nc4 51.Ra6+ Kg7 52.Rc6 Kf7 53.Rc5 Ke6

54.Rxg5 Kf6 55.Rc5 g5 56.Kd4 1-0[/pgn]

[Eelco de Groot]

• Only this evening I saw that Ed devoted a whole thread to this game with notes and all already before Christmas which I did not see. I only skipped through it but Interesting stuff Ed! Maybe we found some things independently.
• Actually 33. Rd6 is a dual that even gets a little bit higher score sometimes than 33. Bd4. Not really an improvement on the text because accuracy of measurement is not high, but still nice. Kaissa can now find it at depth 34 in Multi PV = 6 That is mostly luck, but also nice
  
  [D]rn2r1q1/p5k1/2p2bp1/1p4p1/2P5/4B1PQ/5PK1/3R3R w - -
  
  Engine: Kaissa (Latest version, one thread, 512 MB, Multi PV 6)
  by T. Romstad, M. Costalba, J. Kiiski, G.Linscott
  
  34 25:45 +1.08 33.Bd4 Bxd4 34.Rxd4 Rd8 35.Rxd8 Qxd8
  36.Qe6 Nd7 37.Rd1 Nc5 38.Rxd8 Nxe6
  39.Rxa8 b4 40.Rxa7+ Kf6 41.Kf3 Ke5
  42.Rb7 c5 43.Rb6 Kf5 44.g4+ Ke5
  45.Ke3 Nf4 46.Rc6 (4.049.894.401) 2620
  
  34 25:45 +0.42 33.Rd6 {found!} bxc4 34.Qh6+ Kf7 35.Rxf6+ Kxf6
  36.Bxg5+ Kf7 37.Qh4 Re6 38.Rd1 Qf8
  39.Qxc4 a5 40.Rd3 Qe8 41.Qh4 Kg8
  42.Rd8 Qxd8 43.Bxd8 Nd7 44.Bxa5 Nf6
  45.Bc3 Kf7 46.Qc4 (4.049.894.401) 2620
  
  34 25:45 0.00 33.Qh6+ Kf7 34.Bxg5 Qg7 35.Qh4 Re5
  36.Bxf6 Qxf6 37.Qh7+ Qg7 38.Qh4 Qf6 (4.049.894.401) 2620
  
  34 25:45 0.00 33.cxb5 Re5 34.Bd4 Qd5+ 35.Kg1 Nd7
  36.Qh7+ Kf8 37.Qxg6 Qf7 38.Qxf7+ Kxf7
  39.Rh7+ Ke6 40.bxc6 Nf8 41.Rh6 Rf5
  42.c7 Rc8 43.g4 Rf4 44.Bxf6 Rxf6
  45.Rxf6+ Kxf6 46.Rd8 (4.049.894.401) 2620
  
  34 25:45 -0.50 33.Bc1 a5 34.Qh6+ Kf7 35.Bxg5 Qg7
  36.Qh4 Bxg5 37.Qxg5 Ra7 38.cxb5 Rae7
  39.Rd6 cxb5 40.Qxb5 Kf8 41.Qxa5 Nd7
  42.Rh4 Ne5 43.Rf4+ Rf7 44.Re4 g5
  45.Rd8 Rfe7 46.f4 (4.049.894.401) 2620
  
  34 25:45 -0.53 33.Rd2 a5 34.Qh6+ Kf7 35.Bxg5 Qg7
  36.Qh4 Bxg5 37.Qxg5 Ra7 38.cxb5 Rae7
  39.Rd6 cxb5 40.Qxb5 Kf8 41.Qxa5 Nd7
  42.Rh4 Ne5 43.Rf4+ Rf7 44.Qc5 Kg8
  45.Qd5 g5 46.Re4 (4.049.894.401) 2620
  
  A previous version at high depth gave the following scores:
  
  rn2r1q1/p5k1/2p2bp1/1p4p1/2P5/4B1PQ/5PK1/3R3R w - -
  
  Engine: Kaissa (512 MB)
  by T. Romstad, M. Costalba, J. Kiiski, G. Linscott
  
  44 883:16 +1.62 33.Rd6 Be5 34.Bd4 Bxd4 35.Rxd4 Re7
  36.cxb5 cxb5 37.Rd6 Nd7 38.Rxd7 Rae8
  39.Rd3 Re6 40.Qh6+ Kf7 41.Qxg5 Qg7
  42.Rd7+ R6e7 43.Qd5+ Kf8 44.Qd6 Kg8
  45.Rxe7 Rxe7 46.Qd8+ (137.387.357.691) 2592
  
  44 883:16 +1.50 33.Bd4 Bxd4 34.Rxd4 Re7 35.cxb5 cxb5
  36.Rd6 Nd7 37.Rxd7 Rae8 38.Rd3 Re6
  39.Qh6+ Kf7 40.Qxg5 Qg7 41.Rd7+ R6e7
  42.Qd5+ Kf8 43.Qd6 Kg8 44.Rxe7 Rxe7
  45.Qd8+ Kf7 46.Rh8 (137.387.357.691) 2592
  
  44 883:16 0.00 33.Qh6+ Kf7 34.Bxg5 Qg7 35.Qh4 Re5
  36.Bxf6 Qxf6 37.Qh7+ Qg7 38.Qh4 Qf6 (137.387.357.691) 2592
  
  44 883:16 0.00 33.cxb5 Re5 34.Bd4 Rd5 35.Qh6+ Kf7
  36.Bxf6 Kxf6 37.f4 Nd7 38.bxc6 Nb6
  39.fxg5+ Ke6 40.Qh4 Rc8 41.Rc1 Re5
  42.Rhd1 Ke7 43.Ra1 Re2+ 44.Kg1 Ke8
  45.Rxa7 Qf8 46.Qd4 (137.387.357.691) 2592
  
  44 883:16 0.00 33.Rd2 Re4 34.Rd6 Be7 35.Rd4 Re6
  36.Qh6+ Kf7 37.Bxg5 Qg7 38.Rf4+ Bf6
  39.Bxf6 Qxh6 40.Bg5+ Kg8 41.Rxh6 Nd7
  42.Rfh4 Nf8 43.Rh8+ Kf7 44.Be3 Re5
  45.cxb5 cxb5 46.R4h7+ (137.387.357.691) 2592
  
  43 883:16 0.00 33.Rd3 bxc4 34.Rd6 Be5 35.Rdd1 c5
  36.Qh6+ Kf7 37.Qxg5 Re6 38.Bxc5 Bf6
  39.Qg4 Nc6 40.Rd7+ Ke8 41.Rc7 Ne5
  42.Qe4 Rd8 43.Re1 Ra6 44.Qxe5+ Bxe5
  45.Rxe5+ Re6 46.Re7+ (137.387.357.691) 2592

[Eelco de Groot]

Kaissa 'solving' chess: White has 0.05 advantage (Edit: now it's 0.06)

[D]rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR w KQkq -

Engine: Kaissa (one thread, 512 MB)
by T. Romstad, M. Costalba, J. Kiiski, G. Linscott

18/29 0:02 +0.18 1.d4 Nf6 2.e3 e6 3.c4 d5 4.Nf3 c5
5.Nc3 Nc6 6.Be2 dxc4 7.Bxc4 cxd4
8.exd4 a6 9.d5 exd5 10.Nxd5 Nxd5
11.Bxd5 (3.863.972) 1841

19/28 0:02 +0.28 1.d4 d5 2.c4 e6 3.Nf3 Be7 4.Nc3 Nc6
5.e3 Nf6 6.Bd2 O-O 7.Bd3 Nb4 8.Be2 c5
9.O-O cxd4 10.exd4 (5.418.772) 1841

20/25 0:03 +0.21-- 1.d4 d5 (5.641.533) 1839

20/27 0:03 +0.26 1.d4 d5 2.c4 Nf6 3.Nf3 e6 4.Nc3 Bb4
5.e3 c5 6.Bd3 cxd4 7.exd4 dxc4
8.Bxc4 Qc7 9.Qd3 Nc6 10.O-O O-O
11.Be3 Bd7 12.a3 Be7 (6.762.072) 1819

21/29 0:04 +0.20 1.d4 d5 2.c4 Nf6 3.Nf3 e6 4.Nc3 Bb4
5.e3 c5 6.Bd3 cxd4 7.exd4 dxc4
8.Bxc4 O-O 9.O-O Qc7 10.Qd3 Bd7
11.Bd2 Nc6 12.a3 Bd6 (8.831.642) 1811

22/27 0:06 +0.22 1.d4 d5 2.c4 Nf6 3.Nf3 e6 4.Nc3 Bb4
5.e3 c5 6.Bd3 cxd4 7.exd4 dxc4
8.Bxc4 O-O 9.O-O Bd7 10.Be3 Nc6
11.a3 Bd6 12.Bd3 h6 (11.127.029) 1798

23/27 0:06 +0.14-- 1.d4 d5 (12.142.062) 1798

23/35 0:07 +0.22++ 1.d4 (13.332.761) 1793

23/35 0:07 +0.33++ 1.d4 (13.857.793) 1798

23/35 0:11 +0.22 1.d4 d5 2.c4 Nf6 3.Nf3 e6 4.Nc3 Bb4
5.e3 c5 6.a3 Bxc3+ 7.bxc3 Nc6 8.Bd3 b6
9.cxd5 exd5 10.Ne5 Bd7 11.Nxc6 Bxc6
12.O-O O-O 13.Bb2 Qc8 14.c4 (20.645.425) 1773

24/32 0:12 +0.23 1.d4 d5 2.c4 Nf6 3.Nf3 e6 4.Nc3 Bb4
5.e3 c5 6.Be2 cxd4 7.exd4 dxc4
8.Bxc4 O-O 9.O-O Bd7 10.Qb3 Nc6
11.d5 exd5 12.Nxd5 Nxd5 13.Bxd5 (22.749.916) 1769

25/29 0:14 +0.23 1.d4 d5 2.c4 Nf6 3.Nf3 e6 4.Nc3 Bb4
5.e3 c5 6.Bd2 cxd4 7.exd4 dxc4
8.Bxc4 O-O 9.O-O Nc6 10.Be3 Bd7
11.a3 Bd6 12.Nb5 Bb8 13.Rc1 (26.198.064) 1769

26/32 0:15 +0.31++ 1.d4 (27.874.419) 1771

26/36 0:17 +0.38++ 1.d4 (31.515.019) 1771

26/38 0:23 +0.29 1.d4 d5 2.c4 Nf6 3.cxd5 c6 4.Nf3 cxd5
5.Nc3 Nc6 6.Bf4 Bf5 7.Qb3 Qb6 8.Qxb6 axb6
9.e3 e6 10.Rc1 Ne4 11.a3 Be7 12.Nxe4 Bxe4
13.Nd2 O-O 14.Nxe4 (41.737.796) 1765

27/37 0:31 +0.31 1.d4 d5 2.c4 e6 3.Nf3 Nf6 4.Nc3 dxc4
5.e4 c5 6.Bxc4 cxd4 7.Nxd4 Bc5 8.Be3 O-O
9.O-O Nbd7 10.Be2 a6 11.a3 h6 12.Rc1 Bb6
13.Nf3 Bxe3 14.fxe3 (55.504.085) 1759

28/36 0:36 +0.23-- 1.d4 d5 (64.299.366) 1757

28/36 0:40 +0.16-- 1.d4 d5 (71.496.902) 1758

28/37 0:44 +0.23++ 1.d4 (77.494.791) 1756

28/40 0:55 +0.35 1.d4 d5 2.c4 Nf6 3.cxd5 c6 4.dxc6 Nxc6
5.Nf3 e5 6.e3 exd4 7.exd4 Bb4+ 8.Nc3 Bg4
9.Be2 Bxf3 10.Bxf3 Nxd4 11.O-O Nxf3+
12.Qxf3 O-O 13.Bg5 Qb6 14.Bxf6 (97.915.492) 1757

29/39 1:01 +0.27-- 1.d4 d5 (108.619.602) 1756

29/39 1:07 +0.20-- 1.d4 d5 (118.547.201) 1757

29/41 1:11 +0.27++ 1.d4 (125.578.895) 1760

29/41 1:21 +0.22 1.d4 d5 2.c4 Nf6 3.cxd5 c6 4.dxc6 Nxc6
5.Nf3 e5 6.e3 exd4 7.exd4 Bb4+ 8.Nc3 O-O
9.Be2 Bg4 10.O-O Bxf3 11.Bxf3 Nxd4
12.Bxb7 Rb8 13.Ba6 Qa5 14.Qxd4 (143.562.280) 1760

30/42 1:30 +0.30++ 1.d4 (158.746.514) 1761

30/42 1:39 +0.24 1.d4 d5 2.c4 Nf6 3.Nf3 e6 4.Nc3 Bb4
5.Bg5 Nbd7 6.cxd5 exd5 7.e3 O-O
8.Bd3 h6 9.Bf4 c5 10.O-O Bxc3
11.bxc3 c4 12.Bc2 Re8 13.Ne5 Nxe5
14.Bxe5 (175.231.800) 1760

31/41 2:06 +0.21 1.d4 d5 2.c4 Nf6 3.Nf3 e6 4.Nc3 Bb4
5.Bg5 Nbd7 6.Bxf6 Qxf6 7.Qb3 Bxc3+
8.bxc3 Nb6 9.e3 O-O 10.c5 Nd7 11.Bd3 c6
12.Qc2 h6 13.O-O b6 14.c4 (222.484.582) 1760

32/44 2:56 +0.27 1.d4 d5 2.c4 Nf6 3.cxd5 c6 4.dxc6 Nxc6
5.Nf3 e5 6.e3 exd4 7.exd4 Bb4+ 8.Nc3 O-O
9.Be2 Bg4 10.O-O Bxf3 11.Bxf3 Nxd4
12.Bxb7 Rb8 13.Ba6 Qa5 14.Qxd4 (311.976.990) 1762

33/47 3:41 +0.20-- 1.d4 d5 (391.964.628) 1766

33/48 5:02 +0.12-- 1.d4 d5 (535.394.824) 1768

33/50 8:12 +0.05 1.d4 Nf6 2.c4 e6 3.Nf3 d5 4.e3 c5
5.a3 cxd4 6.exd4 dxc4 7.Bxc4 a6
8.Nc3 b5 9.Bb3 Bb7 10.O-O Be7 11.Re1 O-O
12.Bf4 Nbd7 13.d5 exd5 14.Nxd5 (871.648.883) 1770

34/56 10:14 +0.12++ 1.d4 (1.071.442.314) 1743

34/56 11:32 +0.20++ 1.e4 (1.190.735.521) 1720

34/61 28:06 +0.06 1.d4 Nf6 2.c4 e6 3.Bf4 {!?) d5 4.cxd5 exd5
5.e3 Bd6 6.Bxd6 Qxd6 7.Nf3 Bf5 8.Nh4 Be4
9.Be2 O-O 10.O-O Nbd7 11.Nc3 c6
12.f3 Bg6 13.Nxg6 hxg6 14.e4 (2.740.828.232) 1624

[Eelco de Groot]

• Only this evening I saw that Ed devoted a whole thread to this game with notes and all already before Christmas which I did not see. I only skipped through it but Interesting stuff Ed! Maybe we found some things independently.
• Actually 33. Rd6 is a dual that even gets a little bit higher score sometimes than 33. Bd4. Not really an improvement on the text because accuracy of measurement is not high, but still nice. Kaissa can now find it at depth 34 in Multi PV = 6 That is mostly luck, but also nice

Kaissa needs depth 40 AND Multi PV = 2 now to find 29. Q3. Depth 40 Grrr. Oh well at least it finds it after two hours. But Alpha found it probably much ahead, or at least it saw the potential in the position. Still I'm glad.

[D]rn3r2/p3b1kp/2p5/1p3bp1/4B3/q1P1B1P1/P4P2/3RR1KQ b - -

Engine: Kaissa (one thread, 512 MB)
Version evaluate_011.cpp 29-12-2017 19:54
by T. Romstad, M. Costalba, J. Kiiski, G. Linscott

40 116:04 +0.64 27...Bxe4 28.Qxe4 Kg8 29.Qe6+ Rf7
30.Qg4 Nd7 31.Rxd7 Qxc3 32.Rc1 Qf6
33.Bd4 Qg6 34.Rc7 c5 35.Bxc5 Bf6
36.Be3 Re8 37.Rxf7 Qxf7 38.Rc5 Qxa2
39.Bxg5 Qb1+ 40.Kh2 Qg6 (17.581.186.590) 2524

40 116:04 +0.66 27...Bg6 28.Bxg6 hxg6 29.Qh3 {$1 found! Yay! } Bf6
30.Kg2 Qxa2 31.Rh1 Qg8 32.Rd6 Re8
33.c4 bxc4 34.Qh6+ Kf7 35.Rxf6+ Kxf6
36.Bxg5+ Kf7 37.Qh4 Re6 38.Rd1 Qf8
39.Qxc4 a6 40.Rd3 Qe8 (17.581.186.590) 2524

At depth 39 it was still 0.00:


39 63:06 0.00 27...Bg6 28.Bxg6 hxg6 29.Bd4+ Bf6 30.Bxf6+ Rxf6 31.Re8 Rf8 32.Qe4 Rxe8 33.Qxe8 Qf8 34.Qe5+ Qf6 35.Qc7+ Qf7 36.Qe5+ Qf6 (9.582.268.221) 2530
39 63:06 +0.63 27...Bxe4 28.Qxe4 Kg8 29.Qe6+ Rf7 30.Qg4 Nd7 31.Rxd7 Qxc3 32.Rc1 Qf6 33.Bd4 Qg6 34.Rc7 c5 35.Bxc5 Bf6 36.Be3 Re8 37.Rxf7 Qxf7 38.Rc5 Qxa2 39.Bxg5 Qb1+ 40.Kh2 Qg6 (9.582.268.221) 2530

I almost corrected myself again about 33. Rd6, thinking it was not a dual, because the eval got stuck at +0.49 fot hours, but then the same version found correct continuation but only after 33. Rd6 bxc4:

[D]rn2r1q1/p5k1/2pR1bp1/6p1/2p5/4B1PQ/5PK1/7R w - -

Engine: Kaissa (512 MB)
previous version
by T. Romstad, M. Costalba, J. Kiiski, G. Linscott

44 787:54 +2.06 34.Qh6+ Kf7 35.Rxf6+ Kxf6 36.Qxg5+ {! not Bxg5+} Ke6
37.Rd1 Qf7 38.Qh4 Qf5 39.Bg5 Qd5+
40.Rxd5 cxd5 41.Qg4+ Kd6 42.Bf4+ Kc5
43.Qxg6 Rd8 44.Be3+ Kb4 45.Qf6 Rc8
46.Qd6+ Kb3 47.Bc1 (137.077.984.501) 2899

44 787:54 0.00 34.Qg4 Rxe3 35.fxe3 Qf8 36.Rd2 Qe7
37.Qh3 Qe4+ 38.Kg1 Qxe3+ 39.Rf2 Kf7
40.Kg2 Qe4+ 41.Kh2 Qd4 42.Rxf6+ Qxf6
43.Rf1 a5 44.Qh7+ Ke6 45.Qg8+ Ke7
46.Qh7+ Ke6 (137.077.984.501) 2899

44 787:54 0.00 34.g4 Re4 35.f4 Be7 36.Qf3 Re6
37.Bd4+ Kf7 38.Rxe6 Kxe6 39.Rd1 gxf4
40.Qxf4 Kd7 41.Bf6+ Kc8 42.Bxe7 Qe6
43.Qf8+ Kb7 44.Rb1+ Ka6 45.Ra1+ Kb7
46.Rb1+ (137.077.984.501) 2899

44 787:54 0.00 34.Bc1 Qh8 35.Qxh8+ Rxh8 36.Rxh8 Kxh8
37.Rxf6 a5 38.Rxg6 a4 39.Bb2+ Kh7
40.Rxg5 a3 41.Ba1 Ra7 42.g4 Rd7
43.Rh5+ Kg6 44.f4 Rd2+ 45.Kf3 Nd7
46.f5+ Kf7 47.g5 (137.077.984.501) 2899

44 787:54 -0.08 34.Rb1 c5 35.Bxc5 Nc6 36.Qd7+ Ne7
37.Rxf6 Rad8 38.Qe6 Qxe6 39.Rxe6 Nf5
40.Rb7+ Kh6 41.Rxe8 Rxe8 42.Rc7 g4
43.Ba3 Re4 44.Bc1+ g5 45.Rc5 Kg6
46.Rc6+ Kf7 47.Bxg5 (137.077.984.501) 2899

44 787:54 -0.92 34.Bd4 Bxd4 35.Rxd4 Rd8 36.Rxd8 Qxd8
37.Qh7+ Kf6 38.Qb7 Qd5+ 39.Kh2 c3
40.Qxa8 Qxh1+ 41.Kxh1 c2 42.Qxb8 c1Q+
43.Kg2 Qc5 44.Qd8+ Qe7 45.Qd4+ Kf7
46.Qc4+ Qe6 47.Qa6 (137.077.984.501) 2899

best move: Qh3-h6 time: 787:54.203 min n/s: 2.899.635 nodes: 137.077.984.501

This was only found (by the previous version) at depth 44, more than ten hours at one thread at Multi PV 6 Now Kaissa can find it faster, much faster but none of it is tuned so it might not work at all. I hink it works but of course it still is untuned totally...

[Rebel]

• Only this evening I saw that Ed devoted a whole thread to this game with notes and all already before Christmas which I did not see. I only skipped through it but Interesting stuff Ed! Maybe we found some things independently.

I think that Chris proposed an interesting model how AZ could work.

[shrapnel]

This is the problem with DeepMind not making AlphaZero Commercial, one way or another.
Now every Tom, Dick and Harry will claim his two-bit Engine is a Match for AlphaZero !

[Eelco de Groot]

This is the problem with DeepMind not making AlphaZero Commercial, one way or another.
Now every Tom, Dick and Harry will claim his two-bit Engine is a Match for AlphaZero !

The fact that they so far only published ten games, of which maybe this was one of their best, I don't know, does point out that they not "giving everything away" I mean literally. So there is hope for you

Also maybe you saw the pattern, with every game, Go, chess Shogi, I don't know what else they tried, they just matched the best that was out there, including in some cases their own best effort. It is just a proof of concept, a showcase of what can be achieved. What they actually want to do with it, what they want to achieve will be decided by market forces mostly I suppose. Or they haven't gotten that far yet. Or the effort to see what level actually is possible, in chess for instance, could be exponentially diffcult, like a new NN needs to go up in neurons squared each time. So they better wait with that until the hardware catches up.

[Eelco de Groot]

• Only this evening I saw that Ed devoted a whole thread to this game with notes and all already before Christmas which I did not see. I only skipped through it but Interesting stuff Ed! Maybe we found some things independently.
• Actually 33. Rd6 is a dual that even gets a little bit higher score sometimes than 33. Bd4. Not really an improvement on the text because accuracy of measurement is not high, but still nice. Kaissa can now find it at depth 34 in Multi PV = 6 That is mostly luck, but also nice

Kaissa needs depth 40 AND Multi PV = 2 now to find 29. Q3. Depth 40 Grrr. Oh well at least it finds it after two hours. But Alpha found it probably much ahead, or at least it saw the potential in the position. Still I'm glad.

I think this is the max that is still safe:

evaluate_014

[D]rn3r2/p3b1kp/2p5/1p3bp1/4B3/q1P1B1P1/P4P2/3RR1KQ b - -

Engine: Kaissa (512 MB)
by T. Romstad, M. Costalba, J. Kiiski, G. Linscott

30 3:25 +0.46 27...Bg6 28.Bxg6 hxg6 29.Qh3 Bf6
30.Kg2 Qxa2 31.Rh1 Qg8 32.c4 bxc4
33.Rd6 Be5 34.Re6 Re8 35.Rxe8 Qxe8
36.Qh7+ Kf6 37.Bxg5+ Kxg5 38.f4+ Kf6
39.fxe5+ Qxe5 40.Qh8+ Ke6 (513.740.354) 2496

30 3:25 +0.64 27...Bxe4 28.Qxe4 Kg8 29.Bd4 Rf7
30.Qg4 Na6 31.Re5 Qd6 32.Re6 Qc7
33.Qh5 Bf8 34.Be5 Qc8 35.Rg6+ Rg7
36.Rxg7+ Bxg7 37.Bxg7 Qf5 38.Bd4 Nc7
39.Re1 Ne6 40.Re5 Qf6 (513.740.354) 2496
_________________________________________________

31 4:46 +0.56 27...Bxe4 28.Qxe4 Kg8 29.Bd4 Rf7
30.Qg4 Na6 31.Re5 Qd6 32.Re6 Qc7
33.Qh5 Bf8 34.Be5 Qc8 35.Rg6+ Rg7
36.Rxg7+ Bxg7 37.Bxg7 Qf5 38.Bd4 Nc7
39.Re1 Ne6 40.Re5 Qf6 (718.778.133) 2505

31 4:46 +0.83 27...Bg6 28.Bxg6 hxg6 29.Qh3 Bf6
30.Kg2 Qxa2 31.Rh1 Qg8 32.c4 Re8
33.Rd6 bxc4 34.Qh6+ Kf7 35.Rxf6+ Kxf6
36.Bxg5+ Kf7 37.Qh4 Re6 38.Rd1 Qf8
39.Qxc4 c5 40.Rd8 Qxd8 (718.778.133) 2505
____________________________________________________

32 6:57 +0.35 27...Bg6 28.Bxg6 hxg6 29.Qh3 Bf6
30.Kg2 Qxa2 31.Rh1 Qg8 32.Rd6 Be5
33.Rd3 Bf6 34.c4 bxc4 35.Rd6 Re8
36.Qh6+ Kf7 37.Rxf6+ Kxf6 38.Bxg5+ Kf7
39.Qh4 Re6 40.Rd1 Qf8 (1.057.149.780) 2529

32 6:57 +0.39 27...Bxe4 28.Qxe4 Kg8 29.Bd4 Rf7
30.Qg4 Na6 31.Re5 Qd6 32.Re6 Qc7
33.Qh5 Bf8 34.Be5 Qc8 35.Rg6+ Rg7
36.Rxg7+ Bxg7 37.Bxg7 Qf5 38.Bd4 Nc7
39.Re1 Qg6 40.Qg4 Re8 (1.057.149.780) 2529
___________________________________________________

33 13:05 +0.56 27...Bxe4 28.Qxe4 Kg8 29.Bd4 Rf7
30.Qg4 Na6 31.Re5 Qd6 32.Re6 Qd5
33.Rde1 Nc7 34.R1e5 Qxe6 35.Rxe6 Nxe6
36.Qxe6 c5 37.Be3 Bf6 38.Bxg5 Bxc3
39.Be3 c4 40.a4 bxa4 (1.996.173.557) 2541

33 13:05 +1.02 27...Kg8 {27...Bg6 dropped below this} 28.Bxf5 Rxf5 29.Qh3 Rf7
30.Bd4 Nd7 31.Qxd7 Raf8 32.Qxc6 Qxa2
33.Re6 Rd8 34.Rxe7 Rxe7 35.Qf6 Qe2
36.Rc1 Rf8 37.Qxg5+ Kf7 38.Ra1 Rg8
39.Qd5+ Qe6 40.Qh5+ Rg6 (1.996.173.557) 2541
___________________________________________________

34 15:21 +0.64 27...Bxe4 28.Qxe4 Kg8 29.Bd4 Rf7
30.Qg4 Na6 31.Re5 Qd6 32.Re6 Qd5
33.Rde1 Nc7 34.R1e5 Qxe6 35.Rxe6 Nxe6
36.Qxe6 c5 37.Be3 Bf6 38.Bxc5 Kg7
39.Bd6 h6 40.f4 Rg8 (2.343.543.738) 2543

34 15:21 +1.04 27...Bg6 28.Bxg6 hxg6 29.Qh3 Bf6
30.Kg2 Qxa2 31.Rh1 Qg8 32.c4 Re8
33.Bd4 Bxd4 34.Rxd4 Rd8 35.Rxd8 Qxd8
36.Qe6 Nd7 37.Rd1 Nf8 38.Rxd8 Nxe6
39.Rxa8 b4 40.Rxa7+ Kf6 (2.343.543.738) 2543

[MikeGL]

rn3r2/p3b1k1/2p3p1/1p4p1/8/q1P1B1P1/P4P2/3RR1KQ w - - 0 29
In this position will your engine find Qh3 which engines say is STILL EQUAL?
I used latest SF and the move was not a choice.Engines like 29Bd4+

This game is amazing to me due to this position:
rn3r2/p3b2p/2p3k1/1p3bp1/8/q1P1B1P1/P4PB1/3RR1KQ b - - 0 26
The Queen is positioned on H1 out of play.This looks like something from science fiction,an Alien game!After going over this game,the whole point was that Black's Queenside was NEVER developed,and here AlphaZero makes the TAL decision to hang the knight for unseen compensation,truly amazing!
rnb2r2/p3bpkp/1ppq3N/6p1/Q7/2P3P1/P4PBP/R1B1R1K1 b - - 0 19
The turning point that I found in the game was 32...Re8 in this position
rn2r1q1/p5k1/2p2bp1/1p4p1/2P5/4B1PQ/5PK1/3R3R w - - 0 33
A truly remarkable game,but does your engine find or like 29.Qh3?

SF8 dev thinks Kg2 and Qh3 are just the same because it would transpose, so gave both moves the same score.
Also, as shown in actual game, Qh3 was indeed followed by Kg2 by AlphaZero.
So maybe Kg2 is just as strong as Qh3.