TalkChess.com

[d]1rb2r1k/3p1pp1/p3p2p/1p2P3/4q3/1N2B1R1/PPP1N1P1/2KR1B2 w - - 0 1

Pafifi wrote: ↑Mon Aug 31, 2020 2:39 pm [d]1rb2r1k/3p1pp1/p3p2p/1p2P3/4q3/1N2B1R1/PPP1N1P1/2KR1B2 w - - 0 1

Not with Nf4 delayed by Nc5 for one move, as it seems.

That's Cfish with Pure NNUE eval of the latest net from here

https://tests.stockfishchess.org/nns

1rb2r1k/3p1pp1/p3p2p/1p2P3/4q3/1N2B1R1/PPP1N1P1/2KR1B2 w - - 0 1

Analysis by CF EXT 270820 x64 POP N:

1. +- (3.23): 1.Nc5 Qf5 2.Nf4 Qxe5 3.Bd3 Kg8 4.Rh1 f5 5.Be2 Rf7 6.Ncd3 Qe4 7.Rxh6 Qxe3+ 8.Rxe3 gxh6 9.Nh5 d6 10.Rg3+ Kh7 11.Ndf4 Bd7 12.Rd3 Rb6 13.a3 d5 14.Re3 Rd6 15.Bd3 d4 16.Re5 Kg8 17.Re1 Bc6 18.Kd2 Bd7 19.Ng6 Be8 20.b3 Bc6 21.Nhf4 Rf6 22.Ne5 Be8 23.Rg1 Kg7 24.Rh1 Bf7 25.Nf3 Be8 26.Ke2 Kf8 27.Kf2 Ke7 28.b4 Rd8 29.Ne2 Bf7 30.Ne5 Kd6

2. +- (2.73): 1.Nf4 Bb7 2.Bd3 Qxe5 3.Rh1 Qd6 4.Nh5 Rbc8 5.Nc5 Rg8 6.Bf4 Qxf4+ 7.Nxf4 Rxc5 8.Nh5 b4 9.Nf6 Rd8 10.b3 a5 11.Kb2 Bc6 12.a3 g5 13.Rf1 bxa3+ 14.Kxa3 Kg7 15.Nh5+ Kf8 16.Rf6 Ke7 17.Rxh6 Rb8 18.Nf6 a4 19.b4 Rc3+ 20.Kb2 Rxd3 21.cxd3 Rxb4+ 22.Ka1 Rb3 23.Ng4 Be4 24.Ne5 d6 25.Nc4 Bxd3 26.Na5 Rb1+ 27.Ka2 Be4 28.Rxg5 d5

Depth 43 after some Forward- Backward, no tbs loaded.

Walter Eigenmann has the position in his NIghtmare II collection, line of solution is same with transposition at 4th and 5th up to 8th move.
And it's in Hard Talkchess 2020 as nr.26.

In the second line
> 2. +- (2.73): 1.Nf4 Bb7 2.Bd3 Qxe5 3.Rh1 Qd6 4.Nh5 Rbc8 5.Nc5

looks like 4...Rbc8 is a mistake because 4.Bf4 is a eval jump to 4.xx (better as 4.Nc5)

Paloma wrote: ↑Tue Sep 01, 2020 1:29 am In the second line
> 2. +- (2.73): 1.Nf4 Bb7 2.Bd3 Qxe5 3.Rh1 Qd6 4.Nh5 Rbc8 5.Nc5

looks like 4...Rbc8 is a mistake because 4.Bf4 is a eval jump to 4.xx (better as 4.Nc5)

You probably mean 5.Bf4 resp. 5.Nc5, but if so, you're right. So the transposition I talked about between 4th and 5th move to Walter Eigenmann's solution line and the output- line was not correct in output- line. Yet with 4...Rg8 it's the same eval again of course, so that's the only one correct move order.

Seems to me anyhow, like with LC0 (even if not as much as for this one kind of NNs,) one has to check output- lines of NNUE- eval for sooner none-best moves along them than with classical eval maybe now and then.