An Engine-Fooling Puzzle

Thanks!

How do engines usually evaluate positions like these?  The Chess.com engine seems to have trouble with it.

It doesn't give a draw evaluation until black gets in a position where no matter where white moves black is in stalemate.  Like for instance after 1....Kh8 2. Bc6 Kg8 3. h7+ Kh8 4. Kg5 Kg7 5. h6+ Kh8 the engine evaluation goes from +21.8 to 0 because no matter where white moves black is in stalemate.

Somehow StockFish appears not to have much trouble evauating this position. It's so unpredictable! 

Your posted position is different than mine.  In your position with the bishop on e6 instead of f3 the Chess.com engine will immediately evaluate draw after 1...Kh8 because no matter where white moves next the h7, g7 and g8 squares will all be covered by white attackers, resulting in stalemate.  With the bishop on f3 instead of e6, after 1...Kh8 white can play say 2. Bc6 allowing 2...Kg8 with an evaluation of +21.8.

I admit I did originally post the same position you posted but only had it up for a few minutes before I realized the engine caught the draw immediately and I then replaced it with what's posted now.  So perhaps that's what you saw, but regardless try running the position with the bishop on f3 instead of e6 and see what happens.  

Thanks.

Never saw your original post but I must have read your mind subconsciously! Of course I knew the outcome before posting it.

Now this is surprising as I would think engines would be able to immediately recognize the impossibility of queening in this situation combined with the impossibility of checkmating with bishops of the same color...

Engines are quite good in seeing what they can do but pretty average in recognizing what they can't do. This is very similar to the inability to recognize a fortress without special "fortress software". A fortress is a space you cannot penetrate. Also in draws, a draw is a position where you cannot checkmate.

According to SF we think "weird". We are given a set of rules - moves and goal logics - and then we refuse to think by them and come up with all kinds of strange heuristics. Chances are that SF doesn't even know that a bishop can't change its square color. It might only know what bishops can do.

For those who want to improve in tactics here is an excellent book:

https://amzn.to/2XCsHGZ

This is a forced mate in eight puzzle which the Chess.com engine struggles with, at least initially.  Now there is an alternate solution in which all black responses are not forced.  White has alternate options in this alternate solution as well.  Could the two different types of mates confuse the engine?

This puzzle is courtesy of William Shinkman from the St. Louis Globe Democrat in 1887. 

SF had solved this one before I blinked my eyes but that may depend on the distribution of time amongst running processes. The load will vary minute to minute and hour to hour.

Note that SF was not designed to find the quickest mate but primarily to win games. It will attempt to maximize evaluation scores, then find any mate and finally - given the opportunity - minimize the distance to mate. Therefore direct mate compositions with precise checkmate requirements are like veggies to a wolf. You should be happy it is willing to consume them at all. Note that there are numerous problem types SF cannot solve like helpmates and selfmates. It is fascinating that everyone is interrogating SF on a task it was never designed to do, while there are excellent problem solving tools like Popeye and WinChloe.

Though it's not as interesting as the main puzzle, the puzzle I posted on page 2 has the most difficult first move of the 3 puzzles I've posted.

The 1st move in the "page 2 puzzle" is great but the whole (compo)position is extremely heavy. I am sure there are lighter settings to show a similar idea!

I agree with you. My intention was to create a puzzle which had loads of possible moves such that the correct move was that which looked the worst initially. Consequentially, it looks a bit clunky...

Still, I believe many people would struggle to find the correct move there even given several minutes to think.

I am sure they do, I did! In the composition domain the common way to create hardship for the solvers is to offer a relatively small number of attractive looking alternatives to the correct solution (tries). The solvers don't stumble because there are many but because they look more attractive then the right move. This is commonly considered a better technique for composing problems. Advantages are: (1) positions are lighter and therefore more appealing (2) it is easier to extend the composition with more moves or more ideas when it is light.

Thanks! I'll keep this in mind when thinking about my next puzzle.

Stockfish 200222 takes 5 seconds (6M nodes) to find Bc1

Stockfish is built different these days...