Add OGS 12435854 huhoho 2d vs RoyalZero 9d 8c67ecdc
Mardak opened this issue · comments
https://online-go.com/game/12435854
This game has RZ dropping from 60% winrate after move 148. Analyzing the game with the net used (8c67ecdc-120), the now best (5d6d9c5b-121), and current best 256x20 (1629cbfd-V5) for move 150 with 32000 visits seems to show the live eval thought the win rate of 48% for playing K2 was not actually as good.
weight: 8c67ecdc, visit: 32000
L2 -> 11132 (V: 43.14%) (N: 1.53%) PV: L2 L4 K5 H2 K2 P5 P4 P6 Q7 O3 N6 L6 J7 K6 O7 J5 B10 L12 C11 D11 G17 M11 H10 O12 S15 S16 T16 T17 T15 S18 F18
Q1 -> 6024 (V: 49.93%) (N: 0.03%) PV: Q1 R2 P1 R1 L4 P7 O8 L2 M2 H2 K2 L3 K5 K4 K6 J5 L6 J7 L1 K1 J1 H1 M1 F1 B10
K2 -> 4656 (V: 42.85%) (N: 10.96%) PV: K2 L3 Q1 R2 P1 S2 L4 L2 M2 H2 L1 H1 E1 K1 K4 J3 D1 A5 K5 J5
H2 -> 3323 (V: 43.02%) (N: 3.22%) PV: H2 L4 L3 P6 P4 O8 P7 Q6 R6 Q7 R2 S2 R3 S3 O3 R1 P1 Q2 S15 S16
L4 -> 2985 (V: 40.35%) (N: 64.84%) PV: L4 L2 M2 H2 K2 L3 L1 H1 E1 K1 K5 K4 K6 J5 L6 J7
R2 -> 1495 (V: 43.06%) (N: 1.12%) PV: R2 S2 L4 L2 M2 H2 K2 L3 L1 H1 R3 S3 E1 K1 K4 J3
weight: 5d6d9c5b, visit: 32000
L4 -> 7914 (V: 40.18%) (N: 64.82%) PV: L4 L2 R2 R3 M2 P5 P4 P6 R6 O3 S2 P1 T3 T4 N4 R1 S1 T2 Q1 Q2 O1 S3 K2
K2 -> 6567 (V: 41.13%) (N: 10.26%) PV: K2 L3 Q1 R2 P1 S2 L4 L2 M2 H2 K5 K4 K6 J5 L6 J7 L1 K1 J1 H1 K1 F1 B10
R2 -> 3778 (V: 41.31%) (N: 1.16%) PV: R2 S2 R3 S3 L4 L2 M2 H2 K5 K4 K2 L3 L6 K6 L1 P7 O8 R6 Q6 Q7 P5 P9 S7
L2 -> 3741 (V: 41.29%) (N: 1.71%) PV: L2 L4 K5 P5 P4 P6 R6 O3 H2 J5 G17 F18 G16 H17 S15 S16 T16 T17 T15 S18 B10 C11 A6
R3 -> 3531 (V: 41.34%) (N: 0.28%) PV: R3 S3 L4 L2 R2 S2 M2 H2 K2 L3 K5 J5 L1 K1 J1 H1 K6 J7 M1 F1 S15 S16
H2 -> 3456 (V: 41.24%) (N: 2.83%) PV: H2 L4 K5 P7 O8 J5 K6 L6 J7 M5 N6 N4 P4 O3 R3 S3 R2 M2 S2 R6 Q6 Q7 S7 P5 T6 L8 S15 S16
weight: 1629cbfd, visit: 32000
R2 -> 7510 (V: 42.65%) (N: 2.40%) PV: R2 S2 Q1 R3 P1 S1 L4 L2 M2 H2 K2 L3 L1 K1 K5 K4 J1 H1 K6 J5 L6 J7
L4 -> 5970 (V: 41.24%) (N: 60.58%) PV: L4 L2 R2 S2 R3 S3 Q1 Q2 K2 L3 H2 M2 O3 N1 N4 O1 P4 R1 G17 F18 G16 H17 S15
P1 -> 4986 (V: 42.69%) (N: 0.04%) PV: P1 Q1 L4 L2 M2 H2 K2 L3 K5 J5 L1 H1 E1 K1 J1 J3 M1 K1 G17
H2 -> 4062 (V: 42.45%) (N: 6.85%) PV: H2 L4 L3 L6 J3 P6 P4 O3 N6 P5 Q7 J7 K5 K4 K2 J5 G17 F18 G16 H17 S15
K2 -> 3912 (V: 42.36%) (N: 8.97%) PV: K2 L3 R2 S2 Q1 R3 P1 S1 L4 L2 M2 H2 L1 K1 K5 K4 J1 H1 K6 J5 L6 J7
R3 -> 2884 (V: 42.68%) (N: 0.11%) PV: R3 S3 L4 L2 R2 S2 Q1 Q2 K2 L3 H2 M2 O3 N1 N4 O1 P4 R1 G17 F18 G16 H17 S15
…
L2 -> 23 (V: 32.23%) (N: 1.68%) PV: L2 L4 K5 H2 K2 J5 G17
I'm not really sure what's the right play here. Or maybe there is no definitive right or wrong play at this point? Notably the 256x20 didn't think very highly of the best move 150 L2 of the first network.
As a sanity check, move 148 E5 is generally good with all 3 networks although not quite the 60% win rate from the live game:
weight: 8c67ecdc, visit: 32000
B10 -> 13554 (V: 51.08%) (N: 0.26%) PV: B10 L12 C11 D11 C12 G16 M11 E5 D12 E14 F15 G14 E12 D14 E11
A6 -> 7745 (V: 50.79%) (N: 0.20%) PV: A6 H2 K3 E5 A4 C1 A2 L12 B10 M11 H10 C11 G17 F18 G16 H17 S15 O12 S17
E5 -> 3387 (V: 50.49%) (N: 7.20%) PV: E5 K3 K2 L3 L4 L2 M2 H2 L1 H1 E1 K1 J1 J3 D1 A5
F7 -> 2702 (V: 50.75%) (N: 0.80%) PV: F7 K3 L4 L2 M2 H2 K2 L3 K5 J5 L1 K1 J1 H1 E1 J3 D1 A5 S15
L6 -> 2086 (V: 47.85%) (N: 43.65%) PV: L6 E5 G4 H2 K3 E7 A6 G7 A4 C1 A2 B3 A3 E1 B10 B9 C9 C10 B11
weight: 5d6d9c5b, visit: 32000
A6 -> 15661 (V: 48.35%) (N: 0.26%) PV: A6 H2 K3 E5 A4 C1 F7 F6 B10 L12 A2 M11 H10 C11 G17 F18 G16 H17 S15 S16 T16 T17
B10 -> 3480 (V: 48.33%) (N: 0.36%) PV: B10 L12 M11 H2 K3 E5 C11 D11 S15 S16 T16 T17 T15 S18 G17 F18 G16
L6 -> 3126 (V: 46.37%) (N: 43.14%) PV: L6 E5 G4 H2 K3 E7 A6 G8 H8 G7 K6 J7 A4 C1 A2
E5 -> 2988 (V: 48.02%) (N: 6.89%) PV: E5 K3 L4 L2 M2 H2 K2 L3 L1 H1 E1 K1 J1 J3 D1 A5 S15
J7 -> 2463 (V: 48.18%) (N: 2.86%) PV: J7 E5 G4 H2 K3 E7 K4 K5 F7 D8 D9 C8 C9 B8 E8 B9
F7 -> 1521 (V: 48.26%) (N: 0.75%) PV: F7 K3 R2 S2 L4 L2 M2 H2 K2 L3 L1 K1 R3 S3 J1 H1 E1 J3
weight: 1629cbfd, visit: 32000
F7 -> 13345 (V: 52.09%) (N: 0.26%) PV: F7 L6 G17 F18 G16 H17 J7 K3 H2 P7 O8 R6 Q6 P5 N6 O7 Q7 P4 L2
G17 -> 4557 (V: 52.07%) (N: 0.52%) PV: G17 L12 M11 H2 K3 E5 B10 C11 S15 S16 T16 T17 T15 S18 H17 H18 J17 J18 K17
B10 -> 4011 (V: 52.08%) (N: 0.12%) PV: B10 L12 M11 H2 K3 E5 G17 M12 N13 N12 O13 C10 B9 B11 S15 S16 T16 T17 T15
A6 -> 3738 (V: 52.08%) (N: 0.02%) PV: A6 L12 A4 C1 M11 H2 K3 E5 A2 M12 E1 A3 B3 B17 B18 A3 R2 S2 B3
E5 -> 2792 (V: 51.97%) (N: 2.25%) PV: E5 K3 K2 L4 L3 P7 O8 P5 P4 R6 P6 Q6 P5 Q7 O7 J3 R2 R3 S2 R8 S8 S7 S9 T2
All 3 networks do have the PV: E5 K3 K2, but comparing the win rates from move 148 (50.49%, 48.02%, 51.97%) to the above move 150 K2 (after black plays K3): 42.85%, 41.13%, 42.36%. So even when playing the expected variation, the win rate drops by about 8% for each network, so it seems like the value head is off… ?
Still need to figure out why the sudden drop. Must be some blind point involved somewhere. Perhaps it thought the black group can be killed but actually not?
No idea how to test this right now.