Took the first 1000 positions (not to start with too big data at once) out of HHdbVII, copied those of them being drawn ones, exported them with pgn2epd to .epd and let SF19 then search for best moves with 10"/pos single threaded. On 4 instances only, that was done in a few minutes, because I set the Extra Depth in Shredder GUI to 2 only, letting the GUI load next position always already if correct best move was found and kept stable for 2 iterations, so the 10" were used fully only in case bm was not found by engine.
Indeed that way I found one more somewhat useful (for engine- testing) fortress- position, I hadn't known so far, at least hadn't stored in my big problem- database and couldn't remember it neither:
[pgn][Event "3.hm 12th Polish Chess Fed ty 2024"]
[Site "; Polish Chess Fed ty#3280 4-"]
[Date "2024.12.04"]
[Round "?"]
[White "Tarasiuk=V"]
[Black "(=3111.66a4e3)"]
[Result "1/2-1/2"]
[Annotator ""]
[SetUp "1"]
[FEN "8/qp6/p1p1p3/1PP2p2/KRP3p1/P1N1k3/1PP5/4B3 w - - 0 1"]
[PlyCount "12"]
[GameId "2362877199570320"]
[EventDate "2024.12.04"]
{(EG#26441); source: Polish Chess Fed ty#3280 4-12-2024} 1. b6 (1. Bg3 $2 Qxc5 2. bxa6 bxa6 3. Bc7 Kf3 4. Bb6 Qe5 5. Nd1 Qe2 6. Rb3+ Kf4 7. Ne3 Ke4) 1... Qb8 2. Rb3 $1 (2. Ka5 $2 Qe5 3. Rb3 Qxc5+) 2... f4 $1 (2... e5 3. Ka5 Qg8 4. Nd5+ $1 Ke2 5. Re3+ Kd1 6. Rxe5 g3 7. Bc3 cxd5 8. c6 Qc8 9. Rxd5+ Kxc2 10. c7) 3. Ka5 $1 (3. Bh4 $2 Qe5 4. Bf6 Qxc5 5. Be7 Qf5 6. Nb5+ Kf2) (3. Nb1+ $2 Ke2 4. Kb4 f3) 3... g3 (3... Qd8 4. Na4+ Ke2 5. Bb4 f3 6. c3 Qxb6+ 7. cxb6 $1 f2 8. Nc5 f1=Q 9. Nxb7) 4. Na4+ $1 (4. Nd5+ $2 Ke2) 4... Ke2 5. Bb4 g2 6. c3 g1=Q 1/2-1/2
[/pgn]
Pity, SF19 with 30 threads and 8G hash "finds" 1.b6 within a few seconds yet too, just with -+ eval and not staying stable on this one best move, that was why it wasn't solved in Shredder- GUI within 10" single threaded neither, but if you have a look at the output, the problem of the position for getting used for engine- testing is to be seen clearly:
Code: Select all
8/qp6/p1p1p3/1PP2p2/KRP3p1/P1N1k3/1PP5/4B3 w - - 0 1
Analysis by Stockfish 19:
1.b6 Db8 2.Tb3 f4 3.Sb1+ Ke2 4.Lh4 Dh8
Schwarz steht deutlich besser.
-/+ (-1.56) Tiefe: 10/19 00:00:00
...
1.bxc6 bxc6 2.Lg3 Dxc5 3.Lc7 Kf3 4.Tb3 e5 5.Ld6 Dxc4+ 6.Ka5 f4 7.Lxe5 Dc5+ 8.Sb5+ Ke2 9.Lxf4 axb5 10.Le5
Schwarz hat entscheidenden Vorteil.
-+ (-3.48) Tiefe: 18/73 00:00:05 191MN
1.b6 Db8 2.Lh4 De5 3.Tb3 f4 4.Sb5+ Ke2 5.Sd6 g3 6.Sxb7 g2 7.Sa5 g1D 8.Le7 f3 9.Ld6 De4 10.b7 f2 11.b8D f1D 12.Tc3 Dg5 13.Sb3 Kf2 14.Ka5 e5 15.Kxa6 Dgf5
Schwarz steht auf Gewinn.
-+ (-3.53) Tiefe: 19/73 00:00:07 252MN
...
1.b6 Db8 2.Tb3 f4 3.Lh4 De5 4.Sb5+ Ke2 5.Sd6 g3 6.Sxb7 g2 7.Le7 g1D 8.Ld6 De4 9.Sa5 f3 10.b7 f2 11.b8D f1D 12.Td3 Da1
Schwarz steht klar auf Gewinn.
-+ (-4.36 --) Tiefe: 23/30 00:00:40 1368MN
1.bxc6
Schwarz steht klar auf Gewinn.
-+ (-4.36 ++) Tiefe: 23/72 00:00:50 1715MN
Crystal9 at three trials always was around 10" stable on 1.b6, then already getting eval down to 0.00 soon, so even worse, one engine would solve for correct reasons, another one maybe would get a solved- count too, but incorrectly so.
All in all, yes, there are still treasures to be discovered that way of "automated" filtering of big databases , and for those, who own HHdb, which is the gold standard for endgame studies anyhow to me still, such ways of searching the very very many composed puzzles for some single ones, that are usable for testing modern engines, isn't fully and only frustrating yet still, but seeing the much work of one's own, having to sit at PC by yourself also, and the much hardware- time it yet would take to find such single positions of engine-testing- interest out of the really big data, I don't think I'll go that way much more often than I already had done so far till now. Even more, seeing the positions you'll find that way, to a big part are already known in engines' problem- databases, test- suites and fora then still anyhow, I don't think that to be a really clever way of automation
