Computer test

White to move-
10-15 years or so ago, this position was given to computers.  At that time all computers played the materialistic bxa5,  while any master could easily see that capturing the Rook was White's only move that actualy loses.  White can move the King perpetually and Black can make no headway, so White can force a draw.  Both the loss and the draw, at that time, were beyond a computer's abilitiy to calculate, so the computers always chose the material. I am curious to learn if the average home computer program today would still take the Rook.

Hello.

Loaded this position in tChess Pro on an iPhone 4 and had it evaluate it. The result was bxa5 after a 16 ply search with an -8,65 evaluation.

Stockfish, on the same phone, suggested bxa5 up until ply 20(give or take a few) but switched to Kc2. Sticking to that until ply 28 where I stopped. The evaluation was -#33 (don't know what the hashsign means? mate?)

Stockfish quickly found that c6 is actually the worst move, with a mate in 9 following.

Apart from that, bxa5 is the last choice, although everything evaluates as real bad for white at level 30. All lines include a rook sac on c5. Are you certain there should not be a black pawn on c6?

Hello.

What is that telling me in layman's language? That a computer will find the problem with axb5 after a full 10 move search?

I'm so sorry. I did forget to put in the pawn on c6! 

revised board:

My Rybka 3 took about a minute before it realised that taking the rook was a bad idea.  Houdini seemed to know pretty much instantly to leave the rook alone!

Crafty took two or three minutes to decide not to capture, but it doesn't realize the game is drawn (evaluation at -12)

Yes, Rybka and Houdini still evaluate the position as a win for black, but they manage to avoid taking the rook.

Chessmaster saw that Kc2 was best but said the game was -13.5 or something

Critter gave Kd2 as the correct move immediately, and an evaluation of -0.05. (This is for the diagram in post #6)

Found myself in a similar position in this game against Andremendoza.

Chess.com computer analysis said I was losing by -2.5 but as we humans can see it's a draw which I eventually attained with 3 fold repetition.

http://www.chess.com/echess/game?id=48590917

I suppose that for a programme to realise that a position is frozen the programming would have to tell it to abandon its standard methods of evaluation if, after some number of moves, its evaluation always lay within specified parameters (that is neither increased nor decreased significantly).

But what number of moves would a sensible programme choose for this?  Not a small number I think.

Anyway it is intriguing to try to compare the thinking done by us to realise that no progress can ever be made and the programmes methods.

Thanks all!

I had remembered this test from the earlier days of computer chess as it was used to demonstrate the limitations of computers (at that time) in even some fairly obvious situations and I think trying to show the problems with relying soley on brute force.  Of course algarithms, processors and databases have improved exponentially since then, but I was curious how PCs might handle this test today.  It's curious that while most programs seem to understand very quickly that taking the Rook is losing, they can't seem to predict that not taking the Rook is a draw.  Would that be a fair conclusion?

bxa5 didn't even show up on my Houdini for a split second (interesting).

Very very rarely have I seen an engine evaluate a fortress position correctly in terms of the numerical evaluation.  This is one reason you really have to help the engine through an endgame (or simply ask a stronger player).

Yes, Batgirl, it is.  Programmes have improved because with the exponential increase in memory has come the ability for computers to number crunch, not because programmers have found ways to build in deep understanding of chess.

The reason a modern programme can establish that taking the rook is losing is because if you continue to number crunch up to the required depth the evaluation process starts to give a bigger minus.  But however deep the programme is asked to go it will never evaluate black's extra material as valueless.

Thanks...

Yet a highly talented human, who would probably still lose to the computer most of the time, can evalutate the position very accurately and rather quickly.

Isn't that fascinating?

My free Houdini on a non-special lap, sees that the Rook snatch leads to forced mate in 17 for Black almost instantly. Any other move just leads to an eventual draw, despite Black's massive material advantage. 'Puters & programs have improved, but your point is certainly understood. 'Umans & 'Puters each "think" differently, and each has its best applications. People are always better  over dinner & drinks, though. 

Grobzilla - does your Houdini tell you that the other moves lead to a draw?

I ran it through houdini and it was instantly -13 and played on auto, white resigned in a few moves, all moves being waiting moves.  I doesnt seem to know its a draw because moves dont have to repeat and must consider less material to be a loss after not much change.