Still an issue with pass code? about leela-zero

[sethtroisi]

Another instance

http://eidogo.com/#36738aeOv:0,684

[gcp]

Sabak indeed won't give the correct order of moves, and the network requires a correct history. Also, you should add the "-d" option to disable some passing heuristics (autogtp does this).

I get this output:

./leelaz -d -w 9k.txt
Leela: loadsgf cap2.sgf 612
Leela: genmove w
Playouts: 907, Win: 99.97%, PV: S4 pass pass 
Playouts: 1863, Win: 99.98%, PV: S4 pass pass 
Playouts: 2803, Win: 99.99%, PV: S4 pass pass 
Playouts: 3696, Win: 99.99%, PV: S4 pass pass 
Playouts: 4627, Win: 99.99%, PV: S4 pass pass 
Playouts: 5579, Win: 99.99%, PV: S4 pass pass 
Playouts: 6486, Win: 99.98%, PV: S4 pass pass 
Playouts: 7377, Win: 99.87%, PV: S4 pass pass 
Playouts: 8308, Win: 99.50%, PV: S4 pass pass 
Playouts: 9297, Win: 99.30%, PV: S4 pass pass 
Playouts: 10234, Win: 99.22%, PV: S4 pass pass 
Playouts: 11172, Win: 99.18%, PV: pass 
Playouts: 12117, Win: 99.16%, PV: pass 
Playouts: 13079, Win: 99.16%, PV: pass 
Playouts: 14006, Win: 99.16%, PV: pass 
Playouts: 14958, Win: 99.16%, PV: pass 
Playouts: 15895, Win: 99.18%, PV: pass 
Playouts: 16873, Win: 99.19%, PV: pass 
Playouts: 17825, Win: 99.20%, PV: pass 

pass ->   10466 (V: 100.00%) (N:  8.2%) PV: pass 
  S4 ->    3765 (V: 98.03%) (N: 22.4%) PV: S4 pass pass 
 Q10 ->    2927 (V: 97.54%) (N: 21.2%) PV: Q10 pass pass 
 F19 ->     973 (V: 100.00%) (N:  0.8%) PV: F19 H19 C16 C19 
  E2 ->     365 (V: 99.99%) (N:  0.3%) PV: E2 H4 H3 C1 
18496 visits, 412666 nodes, 18671 playouts, 373 n/s

So I can reproduce the game as it was played (it wants to play S4 early on). It needs more time to see that passing instantly is better.

If we ask black to move after S4:

NN eval=0.000000
Playouts: 970, Win:  4.66%, PV: pass pass 
Playouts: 1930, Win:  7.52%, PV: pass pass 
Playouts: 2874, Win:  9.48%, PV: Q10 E10 T10 B13 
Playouts: 3769, Win: 10.84%, PV: Q10 E10 T10 B13 
Playouts: 4681, Win: 11.94%, PV: Q10 E10 F12 E8 
Playouts: 5549, Win: 12.78%, PV: Q10 E10 F12 E8 
Playouts: 6436, Win: 13.52%, PV: Q10 N10 S7 M3 
Playouts: 7183, Win: 14.06%, PV: Q10 N10 S7 M3 
Playouts: 7905, Win: 14.48%, PV: Q10 N10 S7 M3 
Playouts: 8790, Win: 14.96%, PV: Q10 N10 S7 M3 
Playouts: 9674, Win: 15.35%, PV: Q10 N10 S7 M3 
Playouts: 10524, Win: 15.67%, PV: Q10 N10 S7 M3 
Playouts: 11403, Win: 15.99%, PV: Q10 E10 F12 E8 
Playouts: 12276, Win: 16.26%, PV: Q10 E10 F12 E8 
Playouts: 13114, Win: 16.50%, PV: Q10 E10 F12 E8 
Playouts: 13993, Win: 16.75%, PV: Q10 E10 F12 E8 
Playouts: 14827, Win: 16.50%, PV: Q10 E10 T10 M13 R12 
Playouts: 15567, Win: 15.88%, PV: Q10 E10 T10 B9 T1 
Playouts: 16212, Win: 15.37%, PV: Q10 E10 T10 N2 R12 

 Q10 ->   17026 (V: 17.89%) (N:  0.3%) PV: Q10 E10 T10 N2 R12 
pass ->    3164 (V:  0.00%) (N:  8.3%) PV: pass pass 
20191 visits, 4076541 nodes, 16592 playouts, 331 n/s

It takes a while to see that capturing the large group is good but it thinks it's still losing after that.

So it's working correctly. The network is just dumb.

[killerducky]

I'm able to reproduce it now also. Thank you.