mmt Vs. Ovyron (G4 D5 BG2)

[Ovyron]

It had better not be true of Leela and KMCTS

By chess engines I only meant A/B ones. I don't consider Neural Networks or Monte Carlo Tree Searchers "engines".

[Uri Blass]

Yes, nothing is special about Stockfish regarding mate announcements. Unless there's a bug, mate announcements are maximal (opponent can't delay it further, when announced it's real).

You can expect engines to have bugs in mate announcements because they lose 0 rating points for claiming mate with smaller number of moves and if the engine still win the game inspite of the wrong claim and the author care only about rating then the author has no motivation to find the bug and fix it.

[zullil]

You're just too damn stubborn to admit your wrong!

So my hardware is lacking and way too slow to solve a fastest mate in reasonable time. Your analysis methods are lacking and way too primitive to draw 1.g4 against me despite my hardware?! I guess we'll see, because solving a fastest mate has nothing to do with drawing or winning chess positions.

1.g4 is lost! Your so stubborn...

1. g4 is surely a draw with correct play. So I'll be rooting for White, whoever that is.

[Zenmastur]

You're just too damn stubborn to admit your wrong!

So my hardware is lacking and way too slow to solve a fastest mate in reasonable time. Your analysis methods are lacking and way too primitive to draw 1.g4 against me despite my hardware?! I guess we'll see, because solving a fastest mate has nothing to do with drawing or winning chess positions.

1.g4 is lost! Your so stubborn...

1. g4 is surely a draw with correct play. So I'll be rooting for White, whoever that is.

It's dead lost! No amount of analysis can fix the structural weaknesses that white has inflicted on himself. They will persist right into the endgame. But, If you believe so strongly that you can save the game I'm sure Ovyron will be happy to play you a game!

Regards,

Zenmastur

[jp]

You can expect engines to have bugs in mate announcements because they lose 0 rating points for claiming mate with smaller number of moves and if the engine still win the game inspite of the wrong claim and the author care only about rating then the author has no motivation to find the bug and fix it.

Maybe false mate announcements (i.e. you cannot mate within the number of moves announced) should be given losses. It's said that historically this happened in human's games. I guess that would not work, because the authors would then just turn off the mate announcement so it just shows +150 or something like that.

Apart from bugs, we don't even know that all these authors even try to make their engines search a whole mate tree before announcing. A decade ago people posted on this BB about mate announcements in ~150 or 250 moves, even over a decade before then. There's no way even on current hardware an engine can search all defensive moves to depths 300 or 500. The only possible catch is the TBs. If the announcement was 20 moves engine + 230 moves TB, I guess that might not rule out exhaustive defensive search.

[zullil]

I didn't keep it, it was saved to the engine's learn file so from the distance the engine would know to prune the line and speed up the solving of ...Ke6, and that was it.

It's an anecdote, I don't really care if people believe me or not because I know it happened

Well, there certainly are very simple mate problems that Stockfish still can't handle in reasonable time. For example, Stockfish-dev still struggles with this oldie, which I attribute to Bernhard Bauer from this forum:

[d]rk6/p1r3p1/P3B1Kp/1p2B3/8/8/8/8 w - - 0 1

info depth 51 seldepth 70 multipv 1 score cp -32 nodes 2713568835 nps 2570122 hashfull 1000 tbhits 0 time 1055813 pv e6d5 b8c8 d5a8 c7c4 e5g7 c4a4 a8b7 c8d7 g6h6 b5b4 h6g5 b4b3 b7f3 a4a6 g5f4 d7d6 g7d4 a6a2 f4e3 a7a5 e3d3 a5a4 d3c4 a2c2 c4b4 b3b2 d4b2 c2b2 b4a4 b2b6 f3e2 d6c6 e2d1 c6d5 a4a3 d5c4 d1e2 c4c3 a3a4 b6b4 a4a5 b4b8 e2b5 c3d4 b5e2 d4c5 a5a4 b8b7 e2f3 b7b4 a4a5 b4b1 a5a4 c5d4 f3e2 d4e5 e2d3 b1d1 d3e2 d1e1

With a higher number of threads, SF appears to find this reasonably easily though the depth at which it first finds it seems to vary quite a bit from run to run:
info depth 47 seldepth 17 multipv 1 score mate 5 nodes 984081754 nps 469953082 hashfull 22 tbhits 17322204 time 2094 pv e6d7 b5b4 g6f7 b4b3 f7e8 h6h5 e8d8 g7g6 e5c7

I'm intentionally using default settings. In particular, one thread (deterministic) search. Still can't find mate-in-five ...

info depth 60 seldepth 79 multipv 1 score cp -32 nodes 10948636149 nps 2661592 hashfull 1000 tbhits 0 time 4113566 pv e6d5 b8c8 d5a8 c7c4 e5g7 b5b4 a8d5 c4c5 d5f3 c5a5 g6h6 c8d7 h6g6 a5a6 g6g5 a6a2 g5f4 d7d6 g7f8 d6c7 f8c5 b4b3 c5d4 a7a5 f4e3 c7d6 e3d3 a2h2 d3c4 h2h4 f3d1 a5a4 c4c3 b3b2 d1c2 a4a3 d4g7 d6d5 c3b3 h4h1 g7b2 a3b2 b3b2 d5c4 c2g6 h1h2 b2a3 h2f2 g6e8 c4c3 a3a4 f2e2 e8g6 e2e6 g6f7 e6b6 a4a5 b6b8 f7e6 c3d4 a5a4 b8b6 e6f5 b6b2 a4a5 b2b8 f5h3 d4e3 a5a4 e3f2 h3e6

You're using like 200 threads?

More like 400, yes.

Seems like overkill to solve mate-in-5!

Stockfish-dev (bench 4981770) still needs to reach depth 70 before it finally finds mate-in-5. 85 billion nodes ...

info depth 70 seldepth 88 multipv 1 score mate 5 nodes 85297639755 nps 2752347 hashfull 1000 tbhits 0 time 30990873 pv e6d7 b5b4 g6f7 b4b3 f7e7 b3b2 e7d8 b2b1r e5c7

[jdart]

Mate-in-n is not calculated correctly in many cases because of pruning. If the shortest mating move is a sacrifice for example, but there's another mate that is a bit a longer, the longer mate is likely to be selected because the tree with the sac looks bad compared to the existing mate score. Turning off pruning weakens the engine in general. Stockfish also does aggressive depth reduction in some cases.

[Uri Blass]

Mate-in-n is not calculated correctly in many cases because of pruning. If the shortest mating move is a sacrifice for example, but there's another mate that is a bit a longer, the longer mate is likely to be selected because the tree with the sac looks bad compared to the existing mate score. Turning off pruning weakens the engine in general. Stockfish also does aggressive depth reduction in some cases.

1)Turning off pruning after you find mate when you continue to search does not weaken the engine and stockfish continue to search after finding a mate score.

2)The problem with the mate in 5 problem is not finding a longer mate but not finding even a winning move for a long time.

I think that it is better to have some anti-stockfish rating list that is based on positions when stockfish have trouble with them.
The mate in 5 problem can be one of them when many engines are going to beat stockfish 1.5:0.5 in games based on that position.

[Ovyron]

I think that it is better to have some anti-stockfish rating list that is based on positions when stockfish have trouble with them.

Does Stockfish pick a winning move instead of the mate in 5? If so, it's not incorrect.

In many 2.00 positions that are mate it probably also picks a move that doesn't mate the fastest (including Zenmastur's challenge with +50.00 scores) but we wouldn't say Stockfish has trouble with them, because all moves that mate are equal as long as the game is won.

[zullil]

I think that it is better to have some anti-stockfish rating list that is based on positions when stockfish have trouble with them.

Does Stockfish pick a winning move instead of the mate in 5? If so, it's not incorrect.

In many 2.00 positions that are mate it probably also picks a move that doesn't mate the fastest (including Zenmastur's challenge with +50.00 scores) but we wouldn't say Stockfish has trouble with them, because all moves that mate are equal as long as the game is won.

Here's what Stockfish would do just before finding the mate.

info depth 69 seldepth 91 multipv 1 score cp -23 nodes 73412243114 nps 2745474 hashfull 1000 tbhits 0 time 26739369 pv e6d5 b8c8 d5a8 c7c4 e5g7 b5b4 a8d5 c4c5 d5f3 c8c7 g7h6 b4b3 h6g7 c5a5 f3d1 a5a6 g6f5 a6b6 g7d4 b6b5 f5e4 b3b2 d1c2 a7a5 d4c3 c7b6 c2b1 a5a4 e4d3 b6b7 d3c4 b7a6 c4d4 b5b6 d4c4 b6b3 c3d4 a6b7 d4e5 b7c6 e5d4 c6d7 d4e5 d7e6 e5d4 b3b7 d4c3 b7b8 c3d4 b8b3 d4c3 b3b7 c3d4 a4a3 d4c5 b7c7 c4d4 e6f6 c5a3 c7c1 a3b2 c1b1 b2c3 f6e6 d4e4 b1f1 c3b4 f1h1 e4d4 e6f5 d4c3 h1c1 c3d4 c1d1 d4c5 d1b1 b4c3 b1c1 c5b4 f5e4 c3b2 c1b1