Penrose Puzzle

[Sergei S. Markoff]

I've started Stockfish 8 on Xeon with 20 CPUs and 50 000 Mb Hash. Let see at the morning.

BTW, I think it's worth to try to create NN for defecting such a stuck with 50-move rule positions, I don't think it's really hard to do now. Anyone want to try?) I think all we need is the training dataset, which can be obtained by extracting positions from games of strong engines with high abs(qsearch_score) but ended as draw.

[Dann Corbit]

I've started Stockfish 8 on Xeon with 20 CPUs and 50 000 Mb Hash. Let see at the morning.

BTW, I think it's worth to try to create NN for defecting such a stuck with 50-move rule positions, I don't think it's really hard to do now. Anyone want to try?) I think all we need is the training dataset, which can be obtained by extracting positions from games of strong engines with high abs(qsearch_score) but ended as draw.

A "Tin Can" detection algorithm would solve this in a fraction of a second.

However, I guess it would reduce overall strength of play, and every strong engine will draw this anyway.

The only downside of the big eval is that contests like TCEC would award a win.

[Luis Babboni]

Interesting problem, but the article is just the usual Penrose bullshit on how superior the human mind is, and how this can only be explained by quantum effects.
...

[peter]

Hi Erin!

After some quick Forward- Backward:

8/p7/kpP5/qrp1b3/rpP2b2/pP4b1/P3K3/8 w - - 0 1

Analysis by pedantFishW_2017-03-09_popcnt:

1.Kf1 Bf6 2.Kg2 Be3 3.Kxg3 Be5+ 4.Kh4 Ba1 5.Kh5 Bd2 6.Kg4 Be5 7.Kf3 Bec3 8.Ke4 Bd4 9.Kf5 Bg1 10.Ke5 Be1 11.Ke6 Bg3 12.Ke7 Be5 13.Kd8 Bf6+ 14.Ke8 Bh2 15.Kd7 Bh4 16.Ke6 Bf2 17.Kd7 Bh4
= (0.00) Depth: 89/35 00:01:50 4039MN

[Evert]

A "Tin Can" detection algorithm would solve this in a fraction of a second.

However, I guess it would reduce overall strength of play, and every strong engine will draw this anyway.

Yes.
There is something interesting about being able to recognise fortress positions in general, but in terms of playing strength this is a no-go. Programs may not recognise a fortress, but they will often play them correctly anyway because most moves that break the fortress will lose quickly. The only possible benefit comes from reognising a fortress in he leaves, and steering the game towards it (or away from it). This occasional benefit cannot offset the computational cost.

Doesn't mean such in option isn't interesting, but it would be an analysis mode option geared towards problem solving.

[Evert]

Interesting problem, but the article is just the usual Penrose bullshit on how superior the human mind is, and how this can only be explained by quantum effects.

Aww, I like Penrose when he talks about cosmology.

He's not immune to the affliction that strikes many aging great physicists though: they begin to think they're an expert on everything. Einstein and Bohr suffered from this as well, to an extend.

[hgm]

A "Tin Can" detection algorithm would solve this in a fraction of a second.

However, I guess it would reduce overall strength of play, and every strong engine will draw this anyway.

Yes.
There is something interesting about being able to recognise fortress positions in general, but in terms of playing strength this is a no-go. Programs may not recognise a fortress, but they will often play them correctly anyway because most moves that break the fortress will lose quickly. The only possible benefit comes from reognising a fortress in he leaves, and steering the game towards it (or away from it). This occasional benefit cannot offset the computational cost.

Doesn't mean such in option isn't interesting, but it would be an analysis mode option geared towards problem solving.

The cost of detecting such things should not be so high to have measurable impact on Elo. The clue is that you have multiple pieces with zero mobility, and for programs that evaluate mobility this should be really easy to test. Once it passes the test you can do more expensive stuff.

In this case none of the pieces with zero mobility is blocked by any piece with non-zero mobility, meaning the blocking will be permanent. (If one of the pieces was mobile, all its (pseudo-attacker) would not be permanently trapped, and then their (pseudo-)attackers would not be permanently trapped, etc.) It seems logical to discard such permanently trapped pieces from the material evaluation altogether.

Of course there still is an important distinction between being able to solve such a position when it is presented in the root, or recognize it when it occurs in the leaves. In the root you can do any kind of analysis without having any impact on Elo. This is what humans do, because they are not just capable of playing Chess, but are also expert in taking things apart, and easily recognize where you would have to start, or when it is not possible at all. "The three Bishops won't do it, because they are all on the same color, so you need to liberate oe of the other pieces from its confinement. Oh, too bad, they all block each other. Case solved." Programs for detecting deadlock conditions would "reason" the same, and find the solution instantly. Alpha-beta search, which basically is aimlessly moving your pieces around in the hope something interesting might happen, is completely powerless when othing interesting can happen.

[Uri Blass]

Bad testposition as all 4 king moves draw. But article says: white can even win .

White can win if you add white 4 knights but probably computers are going to be happy with repetition and draw with white when they lose with black.

The plan of white is simple to put the white king at d7 and trade knights at c7 when the last knight mate but you need 4 knights for it if all the bishops on the diagonal so probably 3 are not enough.

[Evert]

Of course there still is an important distinction between being able to solve such a position when it is presented in the root, or recognize it when it occurs in the leaves. In the root you can do any kind of analysis without having any impact on Elo. This is what humans do, because they are not just capable of playing Chess, but are also expert in taking things apart, and easily recognize where you would have to start, or when it is not possible at all. "The three Bishops won't do it, because they are all on the same color, so you need to liberate oe of the other pieces from its confinement. Oh, too bad, they all block each other. Case solved." Programs for detecting deadlock conditions would "reason" the same, and find the solution instantly. Alpha-beta search, which basically is aimlessly moving your pieces around in the hope something interesting might happen, is completely powerless when othing interesting can happen.

Yes, so as fortresses go this isn't a very interesting example either. If black could still shuffle pieces around inside the prison, it would start to get more complicated.
Drawing fortresses where the attacker cannot make progress if the defender keeps his pieces in place are more interesting (of course, examples with few pieces are "solved" by tablebases).

I guess the algorithm would go something along the lines of "if I can make an infinite number of legal quiet moves, can I then manoeuvre my pieces in such a way that I can mate the enemy king?" If not, the question is "can I sacrifice something to break up the enemy position?" If the answer is "yes", that still doesn't proof you can win though.

[Evert]

The plan of white is simple to put the white king at d7 and trade knights at c7 when the last knight mate.

I believe black cannot prevent it.

I'm not convinced it's so easy. If black can trade/give away his bishops, he's stale-mated. In the initial position, all white's knights are trapped and will be captured once they move. Not a problem for the first one or two, but you still need to get that last knight into position.