Can computer analysis really understand human behavior?

Question

I just finished a game. I got into a pretty convenient position at move 17 (I play white):

I noticed that in order to checkmate, I only need to remove the black bishop from protecting e7 and then I can get my knight there and checkmate. So, I figured that if I push the pawn on the b file to the c file, there's a good chance my opponent would capture the knight on c3 "for free" and would allow me to checkmate. My plan worked like a charm but the computer analysis tells me that was a blunder.

I do realize that it might not have been the best move if I would have played against a GM. But in this scenarios which we're both around 1400+, this seemed like the most logical move.

So, I wonder how could I really trust an engine if it only plays GM level and doesn't understand the potential move for a human on a specific level.

Here's the PGN for the game with the analysis:

[fen ""]
1.d4 Nf6 2.c4 d5 3.Nf3 dxc4 4.Nc3 e6 5.e4 Bb4 6.Bg5 h6 7.Bxf6 gxf6 { (+0.60 → +1.11) Inaccuracy. A better move was Qxf6. } 8.Bxc4 Nc6 9.O-O O-O 10.Qd3 e5 11.d5 Bg4 { (+0.94 → +2.81) Mistake. The best move was Na5. } 12.Qe3 { (+2.79 → +1.53) Inaccuracy. A better move was dxc6. } 12...Na5 13.Be2 Kh7 14.Nh4 f5 { (+1.48 → +2.08) Inaccuracy. A better move was Bd7. } 15.Nxf5 Qg5 { (+2.31 → +4.92) Blunder. The best move was Bxf5. } 16.Bxg4 Qxg4 { (+5.23 → +9.72) Blunder. The best move was Qxe3. } 17.Qxh6+ Kg8 18.b3 { (+11.64 → +3.66) Blunder. The best move was d6. } 18...Bxc3 { (+4.09 → +318.00) Blunder. The best move was Qg6. } 19.Ne7# 

Answer 1

This response will be very critical of OP's attitude to chess analysis, and I realize that it may even come across as somewhat harsh. However, it is assumed below that OP is interested in improving, and thus I feel that some level of harshness is justified here.

Playing for cheap tricks - because that is exactly what 18.b3 is - is not something a computer would do. You will never find an engine that "understands" the "motivation" behind playing such a move. It calculates variations, and notices that the move is just not good.

However, you seem to be under the erroneous impression that this is not a desireable property for the engine under these circumstances. In the example position you give, of course 18.b3 is a bad move in every respect. It is good if and only if black is careless enough to take on c3. The move doesn't threaten anything, and it only gives black time to defend a lost position. As you have other ways of playing that, unlike 18.b3, won't jeopardize your winning chances, the move is correctly classified by the computer as a mistake.

So basically: you get a winning position, and you don't know how to win. Seeing no way to win, you play a move that isn't very good if black just ignores the knight on c3, hoping that your opponent blunders. You don't try to play the best move in the position. And yet you feel like there is a reason to object when the engine picks up on this? Remember that there is an iron clad rule when playing chess, that should always be followed:

NEVER UNDERESTIMATE YOUR OPPONENT!

I would absolutely not put it past a 1400 player to spot that 18...Bxc3 is a bad move in that position. Furthermore, playing in a way that relies on your opponent falling for 1-move traps, because thinking of better ideas is too much of a bother, is no way to improve. And if you don't want to improve, why does this question even need to be asked? There is no other obvious point in reviewing the games that you play.

If you were in a lost position instead of a winning one, then your doubt would have been justified, as sometimes in such a situation, only the cheap tricks have a chance at succeeding. But if you have a good position, you shouldn't bash the computer for criticizing your way of handling the position. In analysis, you should under no circumstances assume that one side will make bad moves. Analysis is about finding the truth - not about making yourself feel better about your play.

One final note: if you really want to improve using analysis, you shouldn't rely on the computer for analyzing your games. Focus on trying to find the best moves in the position yourself. The computer can never explain the "why" of a position. That is something you need to figure out for yourself most of the time if you don't have the help of a strong human player.

Answer 2

The idea is right but the execution is wrong. The idea being that you noticed that that bishop is absolutely required to remain on that a3-f8 diagonal or it's checkmate for you. Good job seeing that! But trying to win by "trickery" is not really the way to go in chess. Because even lower-rated opponents may not "fall" for it, especially given enough time to think. Look instead at a more threatening/limiting move for that bishop like a3. Black has no good place to which to move the bishop. Black may well make the same mistake and take the knight, but if he/she does not, you have more good follow up moves.

Answer 3

"Can computer analysis really understand human behavior?". My answer is 'Yes, I believe so - as least as well as any person can'. I have studied many chess games over the past 50 years with a view to understanding how natural thought processes occur, so I think that I am qualified to make such a bold statement.

If we assume that intelligence is essentially a matter of information processing - based on knowledge about the world - then the conclusion must be that computer programs will be capable of emulating natural intelligence to a large extent. I am also assuming of course that human beings will be able to comprehend what goes on 'behind the scenes' and I dare say that I have developed some useful insights in that respect, at least in the domain of chess.

In this particular scenario, the knight move to e7 could satisfy a goal of mate, given that the program/person knows about checkmate situations from prior experience or study. Likewise the black bishop at b4 may be 'activated' (i.e. the internal concept relating to it would be), such that it could counter the proposed knight move. That in turn could lead to a sub-goal of moving the bishop away, perhaps via placing some tasty morsel in its path so to speak.

The last paragraph is deliberately tentative, as the actual behaviour depends upon the relevant knowledge available (which will vary from player to player) and the speed of information-processing involved. It is conceivable that a computer could be equipped with knowledge equivalent to that of a poor human player and thus exhibit similar behaviour. Whether it could actually think in linguistic or pictorial ways is another matter but I don't see why not.

Thus my conclusion is that a suitable computer program could also play in the manner described and might also be able to explain its reasoning. Then we might reasonably decided that it also understands human behaviour.

There is a lot more theory that I could write about but I this is not the best place to do so.