David Wu <lightvector at gmail.com> wrote: > Black (X) to move. The screenshot showed that Leela's policy net put about > 96% probability on a and only 1.03% on b. And that even after nearly 1 > million simulations had basically not searched b at all.
Leela does read out ladders to the end, it doesn't rely on policy network alone, but 0.9.0 had some ladder related bugs that might be causing the behavior you're seeing. I tried running this position into a newer beta version (0.9.5) and my result was: MC winrate=0.484375, NN eval=0.515314, score=B+6.6 Nodes: 20017, Win: 53.83% (MC:49.55%/VN:59.05%), PV: H5 J4 H6 F6 H7 D6 F2 M4 E7 E6 G8 D9 D11 F9 H10 F11 F13 Nodes: 42970, Win: 53.37% (MC:50.00%/VN:57.48%), PV: H5 J5 H6 K4 H7 C10 C8 C13 C15 E10 E8 F2 E2 O4 Q11 Nodes: 65577, Win: 53.10% (MC:50.18%/VN:56.66%), PV: H5 J5 H6 K4 H7 C10 C8 C13 C15 E10 D11 D10 E8 F2 B12 B8 C7 Allowing early exit: low: 52.591342% > high: 52.590360% H5 -> 74538 (W: 53.14%) (U: 50.31%) (V: 56.60%: 2586) (N: 86.0%) PV: H5 J5 H6 K4 H7 C10 C8 C13 C15 E10 D11 D10 E8 F2 B12 B8 C7 B7 Q11 -> 381 (W: 44.91%) (U: 46.15%) (V: 43.38%: 12) (N: 8.8%) PV: Q11 H5 R17 Q17 Q18 R10 F2 -> 131 (W: 41.39%) (U: 41.08%) (V: 41.76%: 6) (N: 4.7%) PV: F2 H5 Q11 R10 Q10 ==================================== 74538 visits, score 53.14% (from 53.08%) PV: H5 J5 H6 K4 H7 C10 C8 C13 C15 E10 D11 D10 E8 F2 B12 B8 C7 B7 75050 visits, 83002 nodes, 75050 playouts, 8547 p/s = H5 So it is possible the newer version has a better policy network, or perhaps how the policy network is used depends on a ladder being detected and the issue you saw was a bug that the ladder situation wasn't recognized. Only Gian-Carlo could tell us for sure. I don't know if ladder status is an input to Leela's policy network or not. -Jonathan
_______________________________________________ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
