An Engine-Fooling Puzzle

I've been half a mathematician and an experienced programmer in an early phase of my life but there are many people more expert on current engines than me. There is a club which dedicates themselves to creating problems and compositions which are hard for engines. This club is not easy to locate on the web and I forgot their name and links. This activity is very useful for testing new versions of engines and develop their skills.

Generally, every problem which stays far away from checkmate positions is difficult for engines. Also problems with unexpected twists in later stages. Engines may go to enormous depths but they use strategems to select promising candidate moves. When they are at depth 70 they may still have missed a great sacrifice on move 14. Given sufficient time they will catch up in the end but there are limits to human patience.

Note that endgames studies often present a hard challenge to engines as many have long solution lines and surprises at every stage. It is well known that even some short problems (#6 and #7) are extremely difficult for engines because every move seems illogical and is therefore skipped as a good candidate until the engine worked through all the remaining moves. Couple that to a position where many pieces move freely and you can see the mountain of lines to investigate.

Note that it becomes increasingly hard to fool all the engines. AlphaZero will self-educate and uses neural networks and monte carlo sampling which is quite different from conventional tree searches. If one engine type doesn't get there, chances are the other type will. And the quantum computer hasn't even arrived yet (in chess)!

Very nice! Yes I am excited about quantum computing. By the way, I happen to be something of a mathematician myself!

By the way you should read Rocky64's blog article on "unsolvable" problems:

https://www.chess.com/blog/Rocky64/stockfish-and-another-forced-mate-problem-it-cannot-solve

And there is lots of other quality stuff in his blogs for aspiring composers!

Thanks, @Arisktotle!

Thanks for the info, I'll check it out right now!

@Rocky64: I revisited the Nikitin problem in your blog and just had an interesting idea. Obviously the composer placed the pieces on distant squares because that is what chess composers do. But he made it for humans, not for engines. What if we slightly change the positions of some units (hopefully) without affecting the solution and thus increase the move potential considerably? For instance, change Rc8 to c7 and run the analysis again. If the solution is still correct I would expect Stockfish to have a much harder time finding it as it would need to investigate many more (paradoxical) moves. Is that observable?

It will take approx 10-15 mins for regular Stockfish on 4 cores PC to solve, but speical version of Stockfish forks can solve within 1 mins.

The reason official Stockfish takes longer is due to removal of fortress detection( e.g such kind of special fortress happens 1 in million games and searching those positions usually waste engine resource and weaken the engine)

That's a most interesting piece of analysis though I clearly haven't worked through all the text. Rocky64 confirmed the authors solution which shows that 1 hour of analysis on his machinery is insufficient for a final verdict (in terms of the set of drawing moves, not on the outcome).

This only increases the likehood that a beautiful endgame study starting with 1. ... Nxc3+ is somewhere there. The composers task has then shifted from eliminating 1. ... Nxc3+ to eliminating 1. ...Bc1. I am pretty sure that our Drunk* author intended to make a real endgame study with an unambiguous first move. So he is now faced with the dilemma that often comes across a composition plan. Do I stay on course at all cost or do I grab the opportunity that offered itself? Or both? Whatever he decides, there is more work to be done!

It would be interesting if the Nikitin could be modified to confuse Stockfish more, but I'm afraid that doesn't look likely. Shifting either WR a square or two down the file allows the Rs to double on the h-file and create cooks (in 7 moves), which the Chess.com Stockfish finds instantly!

Too bad. Of course I could suggest to add a white pawn on h4 on top of it, but this cookstopper (if it works) would violate the purity of the whole white army preventing a black check.

Btw, I noticed something interesting about Nikitins problem. The #7 appears not only the fastest checkmate, but also the only way to win! So it would be a perfect endgame study as well. With only chess.com SF it's hard for me to verify that theory though.

Thank you, drmrboss (and Arisktotle) for your detailed analyses! I am not sophisticated enough in engine studies to fully comprehend everything you posted. But alas, I asked and you answered. I will do my best to understand what you've replied.

If you're interested, there's another similar puzzle I created. The idea was for it to be much like the puzzle posted here, only more elegant and complex. What perplexes me, though, is the engines I show the following puzzle to find it to be rather trivial. This is why I hadn't bothered posting it before:

By looking at the position, I am pretty sure those will never come in my game play. 

Not interested in wasting time!

Fair enough lol. I really was only focusing on a position which lead to a draw that an engine couldn't decipher. That the original position was theoretically possible is merely a consequence of its simplicity. At the same time, I understand why someone would prefer such a situation.

It’s a good puzzle, why belittle someone’s work? 

I just spent an hour composing a reply but chess.com discarded it. Sometimes hate this site

Thank you! I am glad you think it's a good puzzle. Apparently chess.com's engine agrees. It marks the first move as "brilliant." My chess engine thought it was obvious though. Looks like I have work to do if I hope to really confuse these engines.

Arisktotle, I know the feeling. My analysis for the first puzzle was initially completely erased...

It seems that Chess.com "times out" when attempting to post a comment/board analysis that takes an hour or more to compose.  It's been like that as long as I've been commenting on this site, about ten years or so.  I don't know what to do about board analysis but when you're writing a comment that takes around an hour it's a good idea to copy your text.  This way when you post it if it didn't go through you can paste it and re-post it immediately.

Definitely! I often did that but things went well for a while and I got lazy. Also, this time it told me I violated guidelines which of course was a total fabrication.