(Sirobaba F.-Chaika I. ICCF 2011) Fruit frustrations

[Eelco de Groot]

First of all, thanks to Ed for Rebel 14! It's cool! But the search from Fruit is sometimes still frustrating. But Ed does not like search so that is the reason why Gideon is not yet included in the package. Rebel's own search is way too much assembler, 20 years, -or no, more, since about 1980 I think Rebel exists?- work would have to be rewritten to C, C++.
At least that is what I understand from what Ed wrote on his site.

I had a little problem getting Rebel registered as UCI, I am surprised nobody else mentioned that? There is a missing .dll on my machine, it needs VCRUNTIME140.dll which I don't have on my system. Apparently not a problem on other Windows 10?
What I did was download just the missing VCRUNTIME140.dll from https://www.dll-files.com/vcruntime140.dll.html and put that in the same directory as Rebel 14. Make sure that you have the 64 bit version for a 64 bit Windows. Now it installs!

Next I tried this position where Growing Fruit unfortunately has an old problem seeing draws when there are not any, it accepts the draws, at first because it thinks otherwise playing White it is lost but White can win here! (Reversed color position)


[fen]r3n1k1/5rpp/q1p3n1/1pPP2P1/1P2P1B1/pQ1R4/P6B/2K2bR1 w - -[/fen]


This is Sirobaba F.-Chaika I. , from Ernst's thread, but then with reversed colors. White plays 1. Rf3!, then if 1...Bc4, Rebel's choice, 2. Be6! pinning the Rook, Rebel Fruit finds the second move but thinks this is a draw for an eternity I tried to change some of the engine options and that sometimes worked but it is just a random effect, the problem lies in Fruit's search. I vaguely remember it has to do with returning hash table results in PV positions. If in a PV you never return a hash result but always keep searching, the draw results in the hash table eventually will be overwritten with deeper searches, wider intervals etc. And the draws get overwritten eventually. That is why Stockfish does that.

Anyway, here is the result from Rebel 14 overnight after having helped with the first move:


[fen]r3n1k1/5rpp/q1p3n1/1pPP2P1/1P2P1B1/pQ3R2/P6B/2K2bR1 b - -[/fen]

Engine: Rebel 14 (256 MB)
gemaakt door Fabien Letouzey, Pawel Koziol, Chris Whittington

17/38 0:01 -4.12 1...Lc4 2.Dc2 Txf3 3.Lxf3 Da7 4.Lg4 cxd5
5.Le6+ Kh8 6.exd5 Pf8 7.Lg4 Lxd5
8.Df5 Lf7 9.Le5 Kg8 (1.837.921) 1049

18/39 0:02 -4.56 1...Lc4 2.Dc2 Txf3 3.Lxf3 Db7 4.Td1 De7
5.Dd2 Ph4 6.Lh1 cxd5 7.Lg3 dxe4
8.Lxh4 De5 (2.699.209) 1053

19/50 0:07 -3.14 1...Lc4 2.Dd1 Te7 3.d6 Txe4 4.d7 Le2
5.Db3+ Kh8 6.Dxa3 Db7 7.Tf7 Txa3
8.dxe8D+ Txe8 9.Txb7 Lxg4 10.Txg4 Txa2 (7.497.750) 1065

20/60 0:25 -2.53 1...Lc4 2.Dd1 Txf3 3.Dxf3 Da7 4.Le6+ Kh8
5.Th1 De7 6.Dh5 Pf8 7.Le5 Lxa2 8.Df7 Dxf7
9.Lxf7 Lc4 10.La1 Pc7 (29.050.350) 1137

21/60 0:41 -1.50 1...Lc4 2.Dd1 Txf3 3.Dxf3 Ta7 4.Le6+ Kh8
5.Le5 cxd5 6.exd5 Te7 7.Td1 Txe6
8.dxe6 Dxe6 9.Lc3 Kg8 10.Td8 De7
11.Tb8 Lxa2 12.c6 Lc4 (47.579.056) 1149

22/64 1:39 -0.58 1...Lc4 2.Dd1 Txf3 3.Dxf3 cxd5
4.exd5 Da7 5.Le6+ Kh8 6.Df7 Dxf7
7.Lxf7 Pe7 8.Te1 Pxd5 9.Txe8+ Txe8
10.Lxe8 Pxb4 11.c6 Pxa2+ 12.Kb1 Pc3+
13.Ka1 b4 14.Le5 Pd5 (117.433.510) 1177

23/68 3:41 -0.43 1...Lc4 2.Dd1 Txf3 3.Dxf3 Db7 4.Le6+ Kh8
5.Th1 De7 6.Df5 Pf8 7.Le5 cxd5
8.exd5 Lxa2 9.Df7 Dxf7 10.Lxf7 Lc4
11.La1 a2 12.Te1 Tb8 13.Lxe8 Lxd5 (265.730.231) 1202

24/77 5:23 -1.28 1...Lc4 2.Dd1 Txf3 3.Dxf3 Db7 4.Le6+ Kh8
5.Th1 De7 6.Df5 Pf8 7.Le5 cxd5
8.exd5 Lxa2 9.Df7 Dxf7 10.Lxf7 Lc4
11.c6 Td8 12.Td1 a2 13.La1 Pc7 14.g6 h5 (392.882.253) 1213

25/77 9:20 0.00 1...Lc4 2.Le6 Lxb3 3.Lxf7+ Kh8
4.Lxg6 hxg6 5.Tf8+ Kh7 6.Lb8 Pf6
7.gxf6 gxf6 8.Th1+ Kg7 9.Tfh8 g5
10.T1h7+ Kg6 11.Th6+ Kg7 12.T6h7+ (690.656.694) 1232

26/77 23:35 0.00 1...Lc4 2.Le6 Lxb3 3.Lxf7+ Kh8
4.Lxg6 hxg6 5.Tf8+ Kh7 6.Lb8 Pf6
7.gxf6 gxf6 8.Th1+ Kg7 9.Tfh8 g5
10.T1h7+ Kg6 11.Th6+ Kg7 12.T6h7+ (1.793.646.945) 1267

27/79 31:56 0.00 1...Lc4 2.Le6 Lxb3 3.Lxf7+ Kh8
4.Lxg6 hxg6 5.Tf8+ Kh7 6.Lb8 Pf6
7.gxf6 gxf6 8.Th1+ Kg7 9.Tfh8 g5
10.T1h7+ Kg6 11.Th6+ Kg7 12.T6h7+ (2.424.151.570) 1264

28/79 45:25 0.00 1...Lc4 2.Le6 Lxb3 3.Lxf7+ Kh8
4.Lxg6 hxg6 5.Tf8+ Kh7 6.Lb8 Pf6
7.gxf6 gxf6 8.Th1+ Kg7 9.Tfh8 g5
10.T1h7+ Kg6 11.Th6+ Kg7 12.T6h7+ (3.454.892.928) 1267

29/88 85:07 0.00 1...Lc4 2.Le6 Lxb3 3.Lxf7+ Kh8
4.Lxg6 hxg6 5.Tf8+ Kh7 6.Lb8 Pf6
7.gxf6 gxf6 8.Th1+ Kg7 9.Tfh8 g5
10.T1h7+ Kg6 11.Th6+ Kg7 12.T6h7+ (6.598.497.071) 1291

30/92 306:10 0.00 1...Lc4 2.Le6 Lxb3 3.Lxf7+ Kh8
4.Lxg6 hxg6 5.Tf8+ Kh7 6.Lb8 Pf6
7.gxf6 gxf6 8.Th1+ Kg7 9.Tfh8 g5
10.T1h7+ Kg6 11.Th6+ Kg7 12.T6h7+ (24.093.084.074) 1311

31/96 576:14 +2.10 1...Lc4 2.Le6 Lxb3 3.Lxf7+ Kh8
4.Lxg6 hxg6 5.Tf8+ Kh7 6.Lb8 Pf6
7.gxf6 gxf6 8.Th1+ Kg7 9.Tfh8 g5
10.T1h7+ Kg6 11.Th6+ Kf7 12.e5 cxd5
13.T8h7+ Kg8 14.Te7 d4 (46.247.004.990) 1337

When you know this is such an old Fruit problem and superior Rebel 14 NNUE evaluation is helpless to correct this on its own, it's rather frustrating to watch...

[peter]

For this position MultiPV would be just necessary to see, if an engine get's the best move (even if other's are winning too) at least into best three lines of output with enough hardware- time.
Miss that feature with all Rebel- ProDeo-Benjamin-engines since quite a while anyhow.
Yet, if you give in the best move and the engine doesn't get a (for Black) winning eval and a corresponding output- line of its own after quite a time of pondering, it's clear too, that there's a blind spot in the search, thanks for extensive trying, Eelco.

E.g. LC0 with one of the latest bigger nets (780486) doesn't get ...Rf6 in a reasonable time neither (stopped after about 90 minutes with Nvidia 3070ti)

2k2Br1/p6b/Pq1r4/1p2p1b1/1Ppp2p1/Q1P3N1/5RPP/R3N1K1 b - - 0 1

Analysis by Lc0 v0.28.2:
(Sorry for german abbr.)

Code: Select all

31...Le3 32.Lxd6 Dxd6 33.Taa2 Tf8 34.Sf1 Lxf2+ 35.Txf2 Txf2 36.Kxf2 Df6+ 37.Kg1 Db6
  =/+  (-0.52)   Tiefe: 7/13   00:00:00  0kN
...
31...Le3 32.Lxd6 Lxf2+ 33.Kxf2 Dxd6 34.Kg1 d3 35.Da5 Dc6 36.Da2 Tf8 37.Dd2 Db6+
  =/+  (-0.40)   Tiefe: 7/15   00:00:00  0kN
31...d3 32.Lxd6 Dxd6 33.Sxd3 Lxd3 34.Da5 Dd5 35.Te1 Ld8 36.Da2
  =/+  (-0.51)   Tiefe: 7/15   00:00:00  0kN
...
31...d3 32.Lxd6 Dxd6 33.Sxd3 cxd3 34.Da2 Le3 35.Kh1 Th8 36.Df7 Dd7 37.Df6 Td8 38.Sf1 Lxf2 39.Dxf2 Dc7 40.Sd2 e4 41.Te1 Te8 42.Sf1
  =  (-0.24)   Tiefe: 14/35   00:00:03  12kN
31...Td7 32.Lc5 Dh6 33.Sf1 g3 34.Sxg3 Le3 35.cxd4 exd4 36.Sf3 c3 37.Sxd4 Lxf2+ 38.Kxf2 Dxh2 39.Dxc3 Tf7+ 40.Sdf5 Txf5+ 41.Sxf5 Txg2+ 42.Kf1 Dh1+ 43.Lg1+ Kd7 44.Td1+ Ke6
  =/+  (-0.46)   Tiefe: 14/35   00:00:03  13kN
...
31...Td7 32.cxd4 exd4 33.Lc5 Dh6 34.Sf1 d3 35.Td1 Te8 36.Da5 Dc6 37.Da1 Lg6 38.Sd2 De6 39.g3 Lh6 40.Sg2 Lg7 41.Sf4 Dc6 42.Da5 Lf7 43.Sf1 c3 44.Txd3 Txd3 45.Sxd3 Lc4 46.Lxa7 Lxd3 47.Lc5 Kb8 48.Se3 Lh6 49.Sc2 Lg5 50.h4 gxh3 51.Kh2 Lf6 52.Db6+ Dxb6 53.Lxb6 Le5 54.Se1 Lh7 55.Kxh3 Le4 56.Te2 Ld6 57.Te3 Th8+ 58.Kg4 c2 59.Sxc2 Lxc2 60.Te6 Lxb4
  -/+  (-1.05)   Tiefe: 23/68   00:05:00  4019kN, tb=28
...
31...Td7 32.cxd4 exd4 33.Lc5 Dh6 34.Sf1 Te8 35.Da5 Dc6 36.Da3 d3 37.Dc3 De6 38.Td1 Ld8 39.Se3 Le4 40.g3 Lc6 41.S1g2 Lf3 42.Te1 Lf6 43.Dc1 c3 44.Sf1 Df7 45.Txe8+ Dxe8 46.h4 Lg7 47.Kh2 De4 48.Dg5 De5 49.Dc1 Lc6 50.Le3 De4 51.Sd2 Dxb4 52.Tf5 Tc7 53.Tf4 De7 54.Df1 cxd2 55.Lxd2 De2 56.Dxe2 dxe2 57.Txg4 Lf3 58.Tg5 Tc2 59.Le1 Ld4 60.Txb5 Tc1 61.Tb4 Lxg2 62.Kxg2 Txe1 63.Kf3 Lb6 64.Tg4
  -/+  (-1.38)   Tiefe: 28/73   00:30:08  29447kN, tb=769
...
31...Td7 32.cxd4 exd4 33.Sf1 d3 34.Lc5 De6 35.Dc3 Te8 36.Td1 Ld8 37.Se3 Le4 38.Sxd3 Lxd3 39.g3 Lc7 40.Te1 Le4 41.Df6 Lf3 42.Dxe6 Txe6 43.Sc2 Txa6 44.Sd4 Kb7 45.Sxb5 Lb6 46.Lxb6 Txb6 47.Te5 Lc6 48.Sa3 Txb4 49.Tc2 Td4 50.Tc1 Ta4 51.Tc3 Ld5 52.Kf2 Kc6 53.Ke3 Kc5 54.Sc2 Td1 55.Tg5 Ta6 56.Ke2 Th1 57.Th5 Ta2 58.Kd2 Tg1 59.Tg5 a5 60.Txg4 a4 61.Tg5 Tb2 62.h4
  -+  (-1.61)   Tiefe: 31/83   01:25:13  75759kN, tb=5496

yet at least after a few minutes after pondering on position after 31...Rf6, eval becomes a winning one (climbing with enough time higher then the one of 31...Rd7 above) and the output- line shows some correct plies in follow up.
BTW for Salai- study LC0 takes about 10 minutes for 1.f4.

[Eelco de Groot]

Hi Peter,

Thanks for trying the positions and all the references! I would try some programming myself, if I had a compiler for Stock(fish) but for that I have my hope pinned on one time reviving my now not so fast anymore i7. I have a very cheap Dell now, refurbished, ca. 2013, one of the first Haswells. I am very pleased with the noise under load etc. but it is only four cores. It set me back only 280 Euros including wireless keyboard and mouse and replaces a 2005 Windows XP system that I mistreated so it finally refused to start up again. They don't make computers like that anymore, unless maybe by Dell...

For this position MultiPV would be just necessary to see, if an engine get's the best move (even if other's are winning too) at least into best three lines of output with enough hardware- time.
Miss that feature with all Rebel- ProDeo-Benjamin-engines since quite a while anyhow.

I think Rebel never had Multi PV? But now that it is UCI maybe Ed can implement some options Rebel had 'exclude move' that would work well instead and is a standard UCI option in the Shredder UCI protocol too, but not implemented in Fruit. At least I could not get Rebel to exclude a move. Old Toga in some of its final incarnations had some form of simple multithreading but well I don't know It did not work super but better than nothing.

It is possible to switch on Nullmove for Rebel 14 without harm to the playing style (I believe), at least I don't see how that could harm, speed it up a bit. Verification search then I would throw in as well.

I made a Rebel Fruitcocktail, just by changing some UCI options in Rebel 14, that seemed to resolve a win but unfortunately the effect is random and not even reproducible with clean hash I fear.

It produced, one time, this

[fen]r3n1k1/5rpp/q1p1B1n1/1pPP2P1/1Pb1P3/pQ3R2/P6B/2K3R1 b - -[/fen]

Engine: Rebel 14 Fruitcocktail III (256 MB)
gemaakt door Fabien Letouzey, Pawel Koziol, Chris Whittington

17/39 0:01 -5.40 2...Lxb3 3.Lxf7+ Kh8 4.Th1 Ph4
5.Txb3 Da7 6.Le6 Pg6 7.d6 Pf8 8.Lf5 Kg8
9.e5 Df7 10.Lc2 (1.814.530) 1075

18/47 0:03 -5.18 2...Lxb3 3.Lxf7+ Kh8 4.Th1 Ph4
5.Txb3 Da7 6.Le6 Pg6 7.d6 Pf8 8.Lf5 Df7
9.Kb1 Dh5 10.e5 De2 (3.717.430) 1130

19/48 0:07 -5.33 2...Lxb3 3.Lxf7+ Kh8 4.Th1 Ph4
5.Txb3 Da7 6.Le6 Pg6 7.d6 Pf8 8.Lf5 Df7
9.Kb1 Dh5 10.e5 Dxg5 11.Tf1 Dg2
12.Ld3 (9.339.414) 1192

20/65 0:57 0.00 2...Lxb3 3.Lxf7+ Kh8 4.Lxg6 Pc7
5.Th3 cxd5 6.Lxh7 Lxa2 7.Tf1 dxe4
8.Lxc7 Le6 9.Lf5+ Kg8 10.Lh7+ Kh8 (74.588.887) 1302

21/65 10:05 0.00 2...Lxb3 3.Lxf7+ Kh8 4.Lxg6 Pd6
5.Txb3 hxg6 6.Lxd6 Kg8 7.Th1 cxd5
8.exd5 Dc8 9.Tbh3 Df5 10.Th8+ Kf7
11.Txa8 Dc2+ 12.Kxc2 (712.812.711) 1177

22/74 12:31 +8.46 2...Lxb3 3.Lxf7+ Kh8 4.Lxg6 Pd6
5.Th3 Pf5 6.exf5 Ta7 7.Txh7+ Kg8
8.f6 Lxa2 9.Te1 Dc8 10.fxg7 Txg7
11.Te8+ Dxe8 12.Txg7+ Kxg7 13.Lxe8 Lxd5
14.Le5+ Kg8 15.Ld7 (918.937.035) 1222

22/76 24:50 +2.04 2...Taa7 3.De3 Da4 4.Tg2 Ph8 5.Tgf2 Dxb4
6.Ld6 Da5 7.Tc2 Db4 8.Tf5 Pg6 9.Tcf2 Pxd6
10.cxd6 Lxa2 11.Txa2 Dc4+ 12.Kb1 Db4+
13.Ka1 (2.044.924.178) 1371

23/78 32:07 +3.56 2...Taa7 3.De3 Tae7 4.Txf7 Txf7
5.Tg2 Ph8 6.Tf2 Da7 7.Dd4 cxd5
8.exd5 De7 9.De5 Kf8 10.Lxf7 Pxf7
11.d6 Db7 12.De3 Dd7 13.Dxa3 Dg4 (2.693.127.787) 1397

24/78 41:21 +4.15 2...Taa7 3.De3 Pf8 4.Lxf7+ Txf7
5.Txf7 Kxf7 6.Df4+ Kg8 7.Tg3 Le2
8.Tg2 Lh5 9.Tf2 Pd7 10.dxc6 Dxc6
11.Df5 h6 12.Dd5+ Dxd5 13.exd5 hxg5
14.c6 Pb6 15.d6 Lg4 (3.519.935.511) 1418

25/78 56:30 +3.47 2...Taa7 3.De3 Da4 4.Tg2 Ph8 5.Tgf2 Dxb4
6.Ld6 Tab7 7.Dxa3 De1+ 8.Kb2 h6
9.Dc3 Dxc3+ 10.Kxc3 cxd5 11.exd5 Pxd6
12.cxd6 Kf8 13.gxh6 gxh6 14.Lxf7 Pxf7
15.Kd4 Ta7 (4.878.101.354) 1438

26/81 75:57 +3.82 2...Taa7 3.De3 Da4 4.Tg2 Ph8 5.Tgf2 Dxb4
6.Ld6 Tab7 7.g6 hxg6 8.Th2 g5 9.Dxa3 De1+
10.Kb2 Le2 11.dxc6 Ta7 12.Tfh3 Dd2+
13.Kb1 Dd1+ 14.Dc1 Ld3+ 15.Kb2 (6.579.318.647) 1443

27/85 109:55 +4.22 2...Taa7 3.De3 Da4 4.Tg2 Ph8 5.Tgf2 Dxb4
6.Ld6 Tab7 7.g6 hxg6 8.Dxa3 De1+
9.Kb2 Le2 10.Dc3 Dxc3+ 11.Txc3 Lg4
12.dxc6 Tbc7 13.Lxf7+ Txf7 14.Txf7 Pxf7
15.Lg3 Pf6 (9.546.566.445) 1447

27/85 113:52 0.00 2...Lxb3 3.Lxf7+ Kh8 4.Lxg6 hxg6
5.Tf8+ Kh7 6.Lb8 Pf6 7.gxf6 gxf6
8.Tf7+ Kg8 9.Txf6 Txb8 10.Tgxg6+ Kh7
11.Th6+ Kg8 12.Thg6+ (9.904.181.795) 1449

28/85 120:22 0.00 2...Lxb3 3.Lxf7+ Kh8 4.Lxg6 hxg6
5.Tf8+ Kh7 6.Lb8 Pf6 7.gxf6 gxf6
8.Tf7+ Kg8 9.Txf6 Txb8 10.Tgxg6+ Kh7
11.Th6+ Kg8 12.Thg6+ (10.481.927.610) 1451

29/85 132:35 0.00 2...Lxb3 3.Lxf7+ Kh8 4.Lxg6 hxg6
5.Tf8+ Kh7 6.Lb8 Pf6 7.gxf6 gxf6
8.Tf7+ Kg8 9.Txf6 Txb8 10.Tgxg6+ Kh7
11.Th6+ Kg8 12.Thg6+ (11.547.437.517) 1451

30/85 156:27 0.00 2...Lxb3 3.Lxf7+ Kh8 4.Lxg6 hxg6
5.Tf8+ Kh7 6.Lb8 Pf6 7.gxf6 gxf6
8.Tf7+ Kg8 9.Txf6 Txb8 10.Tgxg6+ Kh7
11.Th6+ Kg8 12.Thg6+ (13.662.700.367) 1455

beste zet: Lc4xb3 tijd: 181:00.906 min n/s: 1.455.735 CPU 100.0% n/s(1CPU): 1.455.735 nodes: 15.810.600.000

It lost the solution which is maybe worse than not finding it. Well, I can't reproduce it anyway because it was not started from clean hash.

[carldaman]

It's too bad that Ed decided to go with this buggy Fruit search. It could have been much better with decent search, imo.

[Rebel]

It's too bad that Ed decided to go with this buggy Fruit search. It could have been much better with decent search, imo.

The Fruit search isn't buggy, just an average search, like ProDeo. My interest was in the magic of NNUE and what elo it would bring. And the result is beyond my imagination. And surely, there are inactive 3200+ HCE rated engines waiting to be fertilized with NNUE, can you imagine the outcry here?

[peter]

I think Rebel never had Multi PV? But now that it is UCI maybe Ed can implement some options Rebel had 'exclude move' that would work well instead and is a standard UCI option in the Shredder UCI protocol too, but not implemented in Fruit. At least I could not get Rebel to exclude a move. Old Toga in some of its final incarnations had some form of simple multithreading but well I don't know It did not work super but better than nothing.

As for Toga, I have only dkappe's Dark Toga 1.1 installed and it supports MultiPV, it doesn't exclude move.

...Rf6 is quite soon among the best 3, winning line and eval took me too long. This is after 8 minutes with 30 threads of 16x3.5GHz Ryzen, MultiPV=3, depth 28:


[fen]2k2Br1/p6b/Pq1r4/1p2p1b1/1Ppp2p1/Q1P3N1/5RPP/R3N1K1 b - - 0 1[/fen]

Analysis by Dark Toga 1.1:

1. -/+ (-0.87): 31...Rd7 32.Bc5 Qh6 33.cxd4 exd4 34.Nf1 Re8 35.Qa5 Qc6 36.Qa3 d3 37.Qc3 Be4 38.Nd2 Bg6 39.Nf1 Qd5 40.Rd1 Qe6 41.Bd4 Qe4 42.Bc5 d2 43.Nxd2 Qe5 44.Qxe5 Rxe5 45.h4 Bxh4 46.Rf8+ Bd8 47.Bf2 c3 48.Rc1 Rxd2 49.Rxc3+ Kd7

2. =/+ (-0.62): 31...Be3 32.Bxd6 Qxd6 33.Raa2 dxc3 34.Kh1 Bxf2 35.Rxf2 Rd8 36.Rf1 c2 37.Qe3 Qd7 38.Nxc2 Bxc2 39.Qxe5 Bd3 40.Rf8 Rxf8 41.Qc5+ Kd8 42.Qxf8+ Qe8 43.Qd6+ Kc8 44.Qc5+ Kd7 45.Qxb5+ Ke7 46.Qg5+ Kd7 47.Qxg4+ Kd6 48.Nf1 Ke7 49.Qf4 Qb5

3. =/+ (-0.62): 31...Rf6 32.Bc5 Qe6 33.cxd4 Rxf2 34.d5 Ra2 35.Rxa2 Qf6 36.Bxa7 Rf8 37.Rf2 Qd6 38.Rxf8+ Qxf8 39.Kh1 Bd2 40.Bc5 Qf4 41.Kg1 Bc1 42.Qa5 Qd2 43.Qxb5 Qxe1+ 44.Nf1 Bd3 45.Kh1 Qxf1+ 46.Bg1 Be3 47.Qc6+ Kb8 48.Qd6+ Ka7 49.Qd7+ Kxa6 50.Qc6+ Ka7

Very well does Scorpio 3.0.15.1 using 3070ti GPU:

2k2Br1/p6b/Pq1r4/1p2p1b1/1Ppp2p1/Q1P3N1/5RPP/R3N1K1 b - - 0 1

Analysis by Scorpio 3.0.15.1:

Code: Select all

31...d3 32.Bxd6 Qxd6 33.Nxd3 cxd3 34.Qa2 Be3 35.Nf5 Bxf2+ 36.Qxf2 Qc7 37.Nh6 Rd8 38.Nxg4 Qxc3 39.Rd1 Qd4 40.Qxd4 Rxd4 41.Nxe5 Kc7 42.Kf2 Rxb4 43.Nxd3 Rd4 44.Ke3 Re4+ 45.Kf3 Rd4 
  =/+  (-0.59)   Depth: 12/30   00:00:01  14kN
31...d3 32.Bxd6 Qxd6 33.Nxd3 cxd3 34.Qa2 Be3 35.Kh1 Bxf2 36.Qxf2 Rg7 37.Nf1 Rd7 38.Nd2 e4 39.Qe3 Bg6 40.Kg1 Qe5 41.Rf1 Re7 
  =/+  (-0.54)   Depth: 14/32   00:00:01  28kN
31...d3 32.Bxd6 Qxd6 33.Nxd3 cxd3 34.Qa2 Be3 35.Kh1 Bxf2 36.Qxf2 Rg7 37.Nf1 Rd7 38.Nd2 e4 39.Qe3 Bg6 40.Kg1 Rh7 41.g3 Re7 42.Rf1 Qxa6 43.Nb3 
  =/+  (-0.42)   Depth: 14/32   00:00:02  44kN
31...d3 32.Bxd6 Qxd6 33.Nxd3 cxd3 34.Qa2 Be3 35.Kh1 Bxf2 36.Qxf2 Rg7 37.Nf1 Rd7 38.Nd2 e4 39.Qe3 Bg6 40.Kg1 Rh7 41.g3 Re7 42.Rf1 Qxa6 43.Nb3 
  =/+  (-0.46)   Depth: 14/32   00:00:02  58kN
31...d3 32.Bxd6 Qxd6 33.Nxd3 cxd3 34.Qa2 Be3 35.Kh1 Bxf2 36.Qxf2 Rg7 37.Nf1 Rd7 38.Nd2 e4 39.Qe3 Bg6 40.Kg1 Rh7 41.g3 Re7 42.Rf1 Qxa6 43.Nb3 
  =/+  (-0.44)   Depth: 15/34   00:00:03  72kN
31...d3 32.Bxd6 Qxd6 33.Nxd3 cxd3 34.Qa2 Be3 35.Kh1 Bxf2 36.Qxf2 Rg7 37.Nf1 Rd7 38.Nd2 e4 39.Qe3 Bg6 40.Kg1 Rh7 41.g3 Re7 42.Rf1 Qxa6 43.Nb3 
  =/+  (-0.40)   Depth: 16/40   00:00:03  87kN
31...Be3 32.Bxd6 Qxd6 33.Raa2 dxc3 34.Rae2 Qd4 35.Rxe3 Qxe3 36.Nf1 Qd4 37.Qc1 Rf8 38.Ne3 Rxf2 39.Kxf2 Qd2+ 40.Qxd2 cxd2 41.Ke2 dxe1Q+ 42.Kxe1 g3 43.h4 Kd7 44.Nf1 Be4 45.Nxg3 Bxg2 46.h5 Ke6 47.h6 Kf7 48.Nf5 Kg6 49.Nd6 Bc6 50.Nc8 Kxh6 51.Nxa7 
  =/+  (-0.39)   Depth: 17/42   00:00:04  111kN
 ...
31...Be3 32.Bxd6 Qxd6 33.Raa2 d3 34.Nf1 Bb6 35.Qc1 g3 36.Nxg3 e4 37.Nf1 Qe6 38.Qf4 e3 39.Nxe3 Bxe3 40.Qf3 Bxf2+ 41.Qxf2 Be4 42.Qc5+ Bc6 43.Nf3 Qd5 44.Qxd5 Bxd5 45.Kf2 Rf8 46.h4 Bxf3 47.gxf3 Rh8 48.Ke1 Rh6 49.Ra5 Rh5 50.Ra2 Rxh4 51.Rg2 Rh1+ 52.Kd2 Kc7 53.Rg7+ Kc6 54.Rg6+ Kd5 55.Rg5+ Kd6 56.Rxb5 Rh2+ 57.Kd1 Rc2 58.Rc5 Rxc3 59.f4 Rc2 60.f5 c3 61.f6 Ke6 62.b5 Kxf6 
  -+  (-1.67)   Depth: 30/64   00:03:35  7886kN
31...Rf6 32.Bc5 Be3 33.Kh1 Bxf2 34.Bxb6 Bxg3 35.Bxd4 exd4 36.hxg3 Rf1+ 37.Kh2 Bb1 38.Nf3 Rh8+ 39.Nh4 Re8 40.Nf3 gxf3 41.gxf3 Re2+ 42.Kh3 Bf5+ 43.Kh4 Rxf3 44.Rh1 Re4+ 45.Kh5 Re8 46.Rg1 d3 47.Qa1 d2 48.Qd1 Rd3 49.Kh4 Rh8+ 50.Kg5 Rd5 51.Kf6 Rf8+ 52.Kg7 Rfd8 53.g4 Be4 54.Rf1 R5d7+ 
  -+  (-1.68)   Depth: 30/64   00:03:36  7900kN
31...Rf6 32.Bc5 Be3 33.Kh1 Bxf2 34.Bxb6 Bxg3 35.Bxd4 exd4 36.hxg3 Rf1+ 37.Kh2 Bb1 38.Nf3 Rh8+ 39.Nh4 Re8 40.Nf3 gxf3 41.gxf3 Re2+ 42.Kh3 Bf5+ 43.Kh4 Rxf3 44.Rh1 Re4+ 45.Kh5 Re8 46.Rg1 d3 47.Qa1 d2 48.Qd1 Rd3 49.Kh4 Rh8+ 50.Kg5 Rd5 51.Kf6 Rf8+ 52.Kg7 Rfd8 53.g4 Be4 54.g5 R8d7+ 55.Kf6 Rf5+ 56.Ke6 
  -+  (-1.68)   Depth: 30/64   00:03:36  7917kN
31...Rf6 32.Bc5 Be3 33.Kh1 Bxf2 34.Bxb6 Bxg3 35.Bxd4 exd4 36.hxg3 Rf1+ 37.Kh2 Bb1 38.Nf3 Rh8+ 39.Nh4 Re8 40.Nf3 gxf3 41.gxf3 Re2+ 42.Kh3 Bf5+ 43.Kh4 Rxf3 44.Rh1 Re4+ 45.Kh5 Re8 46.Rg1 d3 47.Qa1 d2 48.Qd1 Rd3 49.Kh4 Rh8+ 50.Kg5 Rd5 51.Kf6 Rf8+ 52.Kg7 Rfd8 53.g4 Be4 54.g5 R8d7+ 55.Kf6 Rf5+ 56.Ke6 
  -+  (-1.69)   Depth: 30/64   00:03:37  7934kN
31...Rf6 32.Bc5 Be3 33.Kh1 Bxf2 34.Bxb6 Bxg3 35.Bxd4 exd4 36.hxg3 Rf1+ 37.Kh2 Bb1 38.Nf3 Rh8+ 39.Nh4 Re8 40.Nf3 gxf3 41.gxf3 Re2+ 42.Kh3 Bf5+ 43.Kh4 Rxf3 44.Rh1 Re4+ 45.Kh5 Re8 46.Rg1 d3 47.Qa1 d2 48.Qd1 Rd3 49.Kh4 Rh8+ 50.Kg5 Rd5 51.Kf6 Rf8+ 52.Ke7 Rfd8 53.g4 Be4 54.g5 Bg6 55.Qg4+ Kc7 56.Qf4+ 
  -+  (-1.70)   Depth: 30/64   00:03:37  7953kN
...
31...Rf6 32.Bc5 Be3 33.Kh1 Bxf2 34.Bxb6 Bxg3 35.Bxd4 exd4 36.hxg3 Rf1+ 37.Kh2 Bb1 38.Nf3 Rh8+ 39.Nh4 Re8 40.Nf3 gxf3 41.gxf3 Re2+ 42.Kh3 Bf5+ 43.Kh4 Rxf3 44.Rg1 d3 45.Qa1 d2 46.Qd1 Rff2 47.g4 Be4 48.g5 Bg6 49.Rh1 Re3 50.Kg4 Be4 51.g6 Bf3+ 52.Kf4 Bxd1+ 53.Kxe3 Re2+ 54.Kf3 Re1+ 55.Kf2 Rxh1 56.g7 Rf1+ 57.Kxf1 Bf3 
  -+  (-1.86)   Depth: 30/64   00:06:05  12575kN

[carldaman]

It's too bad that Ed decided to go with this buggy Fruit search. It could have been much better with decent search, imo.

The Fruit search isn't buggy, just an average search, like ProDeo. My interest was in the magic of NNUE and what elo it would bring. And the result is beyond my imagination. And surely, there are inactive 3200+ HCE rated engines waiting to be fertilized with NNUE, can you imagine the outcry here?

Hi Ed,

Thanks for the new Rebel!

I was just going by what Eelco was showing in this thread. On the other hand, the playing style is outstanding, intrepid and adventurous,
looking to create chances, so un-NNUE-like.

You certainly must be congratulated on that!