Lc0 weak spot position from the past

[Max]

[d]4k3/4p3/3pPp2/2pP1Pp1/1pP3Pp/pP4RP/P5BR/7K b - - 0 1

Even without special knowledge, (most) ab-engines avoid hxg3 after a short think.

Lc0 0.22.0 with weights 42854 struggles here, searching 75M nodes.

- Is there a weights file, which doesn't fall into this trap?
- Are there other engines with the same issue?

net42854.png

[zullil]

[d]4k3/4p3/3pPp2/2pP1Pp1/1pP3Pp/pP4RP/P5BR/7K b - - 0 1

Even without special knowledge, (most) ab-engines avoid hxg3 after a short think.

Lc0 0.22.0 with weights 42854 struggles here, searching 75M nodes.

- Is there a weights file, which doesn't fall into this trap?

320x24.J13-swa-410000

HALTED Nodes: 7.0M, N/s: 75.9k, Time: 1m 32s

6.3% Kf8 Bf3 Kg7 Rc2 Kg8 Rc3 Kh7 Rc1 Kg7 Rb1 Kh6 Kg1 Kh7 Kf2 (N: 5.5M, P: 33.41%)
4.9% Kd8 Bf3 Ke8 Kg2 Kf8 Rh1 Kg7 Rg1 Kh6 Rf1 Kh7 Be4 Kh6 (N: 974k, P: 33.94%)
4.5% hxg3 Be4 gxh2 Bg2 Kf8 Be4 Kg7 h4 Kf8 Kxh2 Kg7 hxg5 fxg5 (N: 492k, P: 32.67%)

[Max]

Yep, 320x24.J13-swa-410000 is great. From the beginning hxg3 is the third choice.

[Max]

Looks like the later T40 nets struggle here (i.e. 42850). Weights file 56215 does fine too.

net56215.png