Mates that are difficult for engines

Fun fact engines can't solve some chess problems that only humans can solve which is kinda funny bc they are rated 4k but they can't solve a position

The issue in my case was not computer speed. I downloaded a version of SF15 with NNUE and I've always used the default NNUE enabled.

This is it in Arena v Rybka with Nalimov tablebase (strictly for basic rules, but OK here because the mate is below 50 moves) with NNUE disabled.

It declares the mate and plays it perfectly accurately. I cut the time down to 12 mins instead of 120 for that run, but I've got it running at 120 at the mo and it's also playing perfectly accurately.

So the difference is probably that you have a version with no NNUE (I believe they dropped it in the latest versions).

Actually mine were running NNUE on both computers, i'll give it a try on my slow computer without tablebases once again with NNUE disabled.

With the NNUE disabled Stockfish 15 found a mate in 30 in 4 seconds, then I tried Stockfish 8 and it found mate in 27 in 5 seconds.

The point about failing on most KNNvKP positions still holds I think. Try for example this random mate in 40 against a tablebase if you have one.

The average mate depth in the endgame is around 58 moves (but that's a bit of a crude measure).

It shows a mate in 41 with Stockfish 15 with NNUE disabled after about 1 minute on my slow computer, I don't have any tablebases on this computer, however a quick glance on chess.com's tablebases shows 1.Ndf1 to be the solution.

It's the right move ok, but does your setup play the mate? Mine (NNUE disabled) comes up with the same move (no mate announcement) but loses the plot after a few moves against the tablebase at 40 moves in 12 minutes. As @Arisktotle mentioned earlier SF is prone to lying about mates. At 40 moves in 120 minutes it's playing it accurately so far.

My prediction was based mainly on past tests with SF8 and SF11. It looks like SF15 has a higher ceiling. It was 35 with the pawn in that position for those versions.

What hash table size do you use?

Hash table is 4 GB on my slow comp, I'll check against the tablebases here to see if it can score the victory.

That could explain some of the difference. I'm using only 2GB.

The Syzygy tablebase should be appropriate for this position. (It's not in general.) Don't envy you the task though in a 40 move mate. If my 40 move in 120 completes successfully (looks accurate so far) possibly better to forget it.

Stockfish 15 successfully executed the mate against the tablebases rather quickly

And perfectly accurately. The one I've got running is going the same way. I'll have to find out what the new ceiling is.

hi

Hello, and welcome.

This is such a fascinating thread! The main reason I think computers can't solve puzzles is because humans can create a puzzle with a logical objective, like "ok white will have to triangulate his king 5 times and each cycle will take 15 moves and then the white knights will checkmate..etc (like that mate in 67) but computers just brute force calculate and if the solution lies outside of the calculation horizon it won't find it, unless you have engines designed for mate searches and pattern recognition like that crystal one.

I've tried one run to determine my SF15 ceiling, giving it mate in 50, time 40 moves in 2 hrs repeating, NNUE disabled and 3GB hash against Rybka/Nalimov in Arena. It fails on that, so my ceiling appears to be over 40 and under 50 with the h3 pawn.

I didn't leave it to complete because it's already blown it on move 7 (move 5 with the 50 move rule in effect) and against Nalimov there's no way back from that.

How does your Arena/SF15 cope with that one?

I'm using my fast computer on this on with SF15 playing white without tablebases and NNUE disabled, and I'mj using SF development version dated April25 with NNUE enabled and 3-4-5 man Syzygy tablebases enabled. I accidentally left the "resign" function on and black resigned after the 5th move . I'll try again.

Yes, I've hit the same problem before on this sort of exercise. Very tedious when you kick it off before you go to bed.

I couldn't seem to get Arena to quit resigning for black, so I played it out against the syzygy tablebases and SF15 looks like it is able to mate in 50 from the root position without NNUE and without tablebases.

Here is the screenshot so far:

Good job, I hope you liked it, and keep in mind that engines can't solve it in general.

SF15 solved it, but in the beginning it didn't really show a mate score, but it played the correct moves. It started going up to a +152 score for white around the 5th move and showed a mate in 36 around the 14th move.

Here is the final screenshot: