Forums

# This puzzle is historically wrong!

Sort:
mohshenas wrote:
BishopTakesH7 wrote:
mohshenas wrote:
Graywing13 wrote:
mohshenas wrote:
Graywing13 wrote:

Bc2+

3.... Kxh5

d8=Q

Your engine should show a win for white now.

4.... Kg4

5.Kc6

5.Kc6 or 5.Qf6 ?

I said 5. Kc6.

I'll try to summarize. The puzzle is wrong because your pruning (you know what that means, right?) engine doesn't see the solution, but you admit the engine is wrong. What's your point, exactly? And yes, I've read your "article" (don't be so humble, btw, it's not an article, it's a whole thesis).

And if we're going for definitions, do kindly define what it means for a puzzle to be wrong.

I'll define a loss for you. I'll even go all fancy and conjure up a recursive definition. A position is lost when it's either a checkmate or there exists such a move that all replies to it lead to a lost position. The position in the puzzle is lost for black and the move is 1.Nf6+

In all honesty though, after white gets a queen, even I will probably win the game against an engine.

I thought Graywing13 was playing white?

mohshenas wrote:
tlay80 wrote:
BishopTakesH7 wrote:
mohshenas wrote:

You can't use the engine line since you are against the engines.

State your reason for the loss without engine lines.

Any new idea?

This must be one of the worst arguments I've ever seen on this site. I try to handle arguments on here reasonably and peacefully, but this is just a terrible point.

Yup.  Also, saying you can't look at it with an engine because you're "against engines" just doesn't follow.  Nobody here is "against engines" in some general way, nor does anyone think they aren't hugely helpful tools.  We do, however, think they have limitations and that this puzzle gives a good example of the sort of problems they occasionally have.

Frankly, I wouldn't have thought that's terribly hard to grasp, but I guess some people have trouble with any judgment that isn't stark black or stark white.

As I said in my prior posts, engines are prone to get stuck in local traps. So, the engines' evaluation numbers are not reliable in general.

In this puzzle, the initial evaluation by stockfish and comodo says about -7; however, after the second or the third move it's about +4. How could you trust that?!

then you should be bashing the computers, not the puzzle

"hOw cOuLd yOu tRuSt tHaT?"

that is doubting the engines, not the puzzle, while this whole thread is about how this puzzle is wrong

you just switched you're argument without noticing it

it becomes +4 from -7 since the computer now sees something it didnt before, obviously

I think Dr. mohshenas's idea is that traditional engines like Stockfish and Komodo are unreliable, but he has done a deeper analysis with a hybrid engine of his own devising (an engine supposedly not susceptible to the same errors of analysis), which analysis proves that white is lost in the initial position.

Be that as it may, the good doctor has made at least two fatal errors of logic, and one of supposition. (1) The way to prove a win for black is by taking account all of white's possible moves. However he rejects 1.Nf6+ not by analysis but by asserting his engine says 1.Nxe3 is "best". (2) Since the initial position is a win for white (see point number 3), and his engine asserts a win for black, we can therefore deduce that his novel hybrid engine *is* susceptible to the same types of errors of analysis as other engines. (3) He wrongly supposes that his engine is battling against human assertions. But he fails to take into account that Van Breukelen's puzzle has been analyzed quite extensively by various engines and even by tablebases, so we know with quite high reliability that it *is* in fact a win for white.

yetanotheraoc wrote:

I think Dr. mohshenas's idea is that traditional engines like Stockfish and Komodo are unreliable, but he has done a deeper analysis with a hybrid engine of his own devising (an engine supposedly not susceptible to the same errors of analysis), which analysis proves that white is lost in the initial position.

Be that as it may, the good doctor has made at least two fatal errors of logic, and one of supposition. (1) The way to prove a win for black is by taking account all of white's possible moves. However he rejects 1.Nf6+ not by analysis but by asserting his engine says 1.Nxe3 is "best". (2) Since the initial position is a win for white (see point number 3), and his engine asserts a win for black, we can therefore deduce that his novel hybrid engine *is* susceptible to the same types of errors of analysis as other engines. (3) He wrongly supposes that his engine is battling against human assertions. But he fails to take into account that Van Breukelen's puzzle has been analyzed quite extensively by various engines and even by tablebases, so we know with quite high reliability that it *is* in fact a win for white.

He even fails to understand completely that if the engine prunes the first move of a position, he will not come back to it even if you allow it to "think" for several hours. But if you insert the move by hand and let it think from there, then it's highly probable that it will find the optimal move (in our case: 1.Nf6) very quickly.

Or sometimes not-so-quickly, in our example after forcing 1.Nf6+, it took several minutes at my average laptop (i7-9750H using 10 cores) to see that 2.Nh5+ wins, even though this move was clearly the top choice.

Again for the record, the latest Crystal 4.0, a Stockfish child some 20 days old, with NNUE=off and Clean Search=ON finds 1.Nf6+ in less than 10 seconds and even at very low depth!

It's a little surprising Crystal "found" 1.Nf6+ at Depth=20. Did you step backwards through the winning line, so the PV was already in the cache? Is it a cloud engine on some big iron? And I am not familiar with Clean Search. Does that mean pruning is off? Still hard to explain 10 seconds.

Anyway, as I am sure you know, it's a bit of confirmation bias for a human to say "my engine found the correct move in XX seconds". It's more conservative to say that at time XX the engine considered this move best, without projecting any finality or certainty on this evaluation. If we let the engine run then it may start to think the currently top move is worse than one of the others, or even evaluate it as losing due to some branch or other.

yetanotheraoc wrote:

It's a little surprising Crystal "found" 1.Nf6+ at Depth=20. Did you step backwards through the winning line, so the PV was already in the cache? Is it a cloud engine on some big iron? And I am not familiar with Clean Search. Does that mean pruning is off? Still hard to explain 10 seconds.

Anyway, as I am sure you know, it's a bit of confirmation bias for a human to say "my engine found the correct move in XX seconds". It's more conservative to say that at time XX the engine considered this move best, without projecting any finality or certainty on this evaluation. If we let the engine run then it may start to think the currently top move is worse than one of the others, or even evaluate it as losing due to some branch or other.

I did not forward any moves to it- it started computing from the initial position.

It is a heavily modified Stockfish, with very different pruning behavior (much less aggressive) and special fortress-detecting code. It has a performance handicap due to the added code, but it is very reliable when time is not an issue (e.g. ICCF games).

I think that "clean search" just starts the engine with an empty hash, but you'd rather ask more tech details to the developer, J. Ellis.

https://github.com/jhellis3/Stockfish/tree/crystal

yetanotheraoc wrote:

I think Dr. mohshenas's idea is that traditional engines like Stockfish and Komodo are unreliable, but he has done a deeper analysis with a hybrid engine of his own devising (an engine supposedly not susceptible to the same errors of analysis), which analysis proves that white is lost in the initial position.

Be that as it may, the good doctor has made at least two fatal errors of logic, and one of supposition. (1) The way to prove a win for black is by taking account all of white's possible moves. However he rejects 1.Nf6+ not by analysis but by asserting his engine says 1.Nxe3 is "best". (2) Since the initial position is a win for white (see point number 3), and his engine asserts a win for black, we can therefore deduce that his novel hybrid engine *is* susceptible to the same types of errors of analysis as other engines. (3) He wrongly supposes that his engine is battling against human assertions. But he fails to take into account that Van Breukelen's puzzle has been analyzed quite extensively by various engines and even by tablebases, so we know with quite high reliability that it *is* in fact a win for white.

First of all. I've been doing research on the combinatorial properties of chess positions. So, let me have the right not to trust the stereotypes in chess programming.

You did put forth several good arguments. I would like to discuss them.

1. It is possible that the engine follows line 1.Nf6+ and finds a checkmate in favor of White in "n" moves. However, if it follows 1.Nxe3, it will eventually find the checkmate for White in "n-1" moves. So, it's better to choose line 1.Nxe3.

2. Yes. I totally agree with you. As I described in the first post, my first approach is a simple combination of engines smartly. It is indeed prone to errors. My second approach is not, and it is indeed not an engine. Please read my previous posts.

3. Engines have been continuously improving for decades as new mathematics, and powerful hardwares emerge. However, I'm not sure anyone has found a more-than-eight-move tablebase up to this moment.

mohshenas wrote:
yetanotheraoc wrote:

I think Dr. mohshenas's idea is that traditional engines like Stockfish and Komodo are unreliable, but he has done a deeper analysis with a hybrid engine of his own devising (an engine supposedly not susceptible to the same errors of analysis), which analysis proves that white is lost in the initial position.

Be that as it may, the good doctor has made at least two fatal errors of logic, and one of supposition. (1) The way to prove a win for black is by taking account all of white's possible moves. However he rejects 1.Nf6+ not by analysis but by asserting his engine says 1.Nxe3 is "best". (2) Since the initial position is a win for white (see point number 3), and his engine asserts a win for black, we can therefore deduce that his novel hybrid engine *is* susceptible to the same types of errors of analysis as other engines. (3) He wrongly supposes that his engine is battling against human assertions. But he fails to take into account that Van Breukelen's puzzle has been analyzed quite extensively by various engines and even by tablebases, so we know with quite high reliability that it *is* in fact a win for white.

First of all. I've been doing research on the combinatorial properties of chess positions. So, let me have the right not to trust the stereotypes in chess programming.

You did put forth several good arguments. I would like to discuss them.

1. It is possible that the engine follows line 1.Nf6+ and finds a checkmate in favor of White in "n" moves. However, if it follows 1.Nxe3, it will eventually find the checkmate for White in "n-1" moves. So, it's better to choose line 1.Nxe3.

2. Yes. I totally agree with you. As I described in the first post, my first approach is a simple combination of engines smartly. It is indeed prone to errors. My second approach is not, and it is indeed not an engine. Please read my previous posts.

3. Engines have been continuously improving for decades as new mathematics, and powerful hardwares emerge. However, I'm not sure anyone has found a more-than-eight-move tablebase up to this moment.

Rather than guessing and rattling non sense in the forum with your own/basic knowledge and own conclusion, I suggest to learn some basic about "Alpha Beta pruning" and you will see why Stockfish did not find winning move very quickly. Given enough time like 1-2 hours in decent hardware, many engines solved that puzzles since 10 years ago.

Stockfish is the most aggresively pruned engine and see extreme depth 20-40 within a fraction of second but it  also has to pay the cause like missing some moves.

mohshenas wrote:

1. It is possible that the engine follows line 1.Nf6+ and finds a checkmate in favor of White in "n" moves. However, if it follows 1.Nxe3, it will eventually find the checkmate for White in "n-1" moves. So, it's better to choose line 1.Nxe3.

What you've been smoking while working on your combined search technique?

It surely is pretty evil stuff.

Ok, this used to be amusing, but now it's straight up laughable.

> 1. It is possible that the engine follows line 1.Nf6+ and finds a checkmate in favor of White in "n" moves. However, if it follows 1.Nxe3, it will eventually find the checkmate for White in "n-1" moves. So, it's better to choose line 1.Nxe3.

Sure. Where is this checkmate? Forget "n-1", where is ANY checkmate? After Nxe3 the position is blatantly lost for white, white is clear piece down, black pawns will promote, white pawns will be lost one by one, because he has nothing to protect them with. You don't need an engine (or a science degree) to see that, you just need to know chess rules and be able to count to three.

I would like to thank everyone who participated in this forum. Now, I'm totally convinced that I should discuss my point of view and the issues with the stockfish's and other engines' developers.

I'm also convinced that this puzzle does not serve its designers' purpose.

--

This is my last post in this thread. Have a nice day and stay safe.

mohshenas wrote:

I would like to thank everyone who participated in this forum. Now, I'm totally convinced that I should discuss my point of view and the issues with the stockfish's and other engines' developers.

I'm also convinced that this puzzle does not serve its designers' purpose.

This is my last post in this thread. Have a nice day and stay safe.

Well mohshenas, I can see the validity of some of the points you were making and I can't respond to them in terms of complexity theory which I am insufficient familiar with. I rely on reports of this puzzle having been solved with the help of tablebases which is not impossible by extending 7/8-piece tablebases with "position directed searches" - exclusively targeting this particular puzzle position.

Rather than discussing your ideas with stockfish and lc0 designers, I'd suggest you take them on in a direct confrontation of the computer programs. As both of you predict different outcomes, it should be easy to settle who's right. The proof of the puzzle is in the playing.

Awww, the circus has left the town, what a shame.

From the moment Graywing13's challenge was accepted, eventual "toys ejected from pram" was the foreseen outcome. That's why I mentioned Shakespeare above, the bard was a GM of human nature.