PDA

View Full Version : Does computer miscalculate?



Igor_Goldenberg
09-05-2006, 03:38 PM
I play Fritz frequently (usually with negative result:( ). It calculates moves much better then me. Yet I was amazed on different occasions to witness a move well below 2000 standard.

Once it allowed me to sacrifice a queen for a rook, with Fritz estimation changing from +5 to -5 in a three semi-moves. I guess a long sequence of moves which are essentially forced (but do not appear so) is still a problem for a computer.

qpawn
25-05-2006, 01:49 PM
I find that Fritz makes serious positional errors. For instance lastw eek it snatched a pawn to go one pawn up. However Fritz in doing so took on tripled pawns. Yet Fritz viewed its position as being clearly better!

Desmond
25-05-2006, 02:17 PM
It is hard for a computer to perceive a move as "forced". They generally follow all stems in the search tree. The variations make this tree expand exponentially and it cumbersome for the computer to continue looking deeper. As you have noted, however, this is usually a very effective (if not efficient) process.

Igor_Goldenberg
26-05-2006, 01:58 PM
I recently read an account by Bareev about fated 8th game in Brissago (Kramnik - Leko), when Leko won in Marshall with 26...Qd3!.

According to Bareev, they analysed this variation, and computer estimated this position as won for white. After about two minutes of thinking, it changed estimation from +- to -+.

However, when analysing at home before the game, Kramnik's team could not afford more then 10-15 seconds for each postion, and therefore didn't notice a hole in the variation.

Desmond
26-05-2006, 02:25 PM
However, when analysing at home before the game, Kramnik's team could not afford more then 10-15 seconds for each postion, and therefore didn't notice a hole in the variation.

That strikes me as quite odd. Seems more like blitz analysis.

Kevin Bonham
26-05-2006, 02:45 PM
I find that Fritz makes serious positional errors. For instance lastw eek it snatched a pawn to go one pawn up. However Fritz in doing so took on tripled pawns. Yet Fritz viewed its position as being clearly better!

It's not always an error to go one pawn up at the cost of accepting tripled pawns. In a lot of Winawer positions White has a choice to make on that issue and in some positions accepting the tripled pawn can be positionally crushing quite aside from it winning you a pawn. In many others it is, in the long term, a game-losing error. Would be interesting for me to see the position.

I've sometimes found computers lose because of snatching off a rook pawn with a bishop in a case where the bishop can be hemmed in by an immediate P-N3 and then captured later. The computer won't fall for it if the recapture is forced in two or three moves but may fall for it if there is some preliminary stuff that the opponent has to do before hunting down the trapped bishop with a king, meaning that the recapture is too far in advance for the computer to see it.

Igor_Goldenberg
26-05-2006, 02:47 PM
That strikes me as quite odd. Seems more like blitz analysis.

When analysing an opening, one has to go through a lot of possibilities.

Desmond
26-05-2006, 02:47 PM
It's not always an error to go one pawn up at the cost of accepting tripled pawns. In a lot of Winawer positions White has a choice to make on that issue and in some positions accepting the tripled pawn can be positionally crushing quite aside from it winning you a pawn.

yes, I have to agree with Kevin there. If black can't win one or two of those Winawer C-pawns more or less immediately they become crushing.

Desmond
26-05-2006, 02:51 PM
When analysing an opening, one has to go through a lot of possibilities.

Of course. However, if you consider that the team would have analysed many almost identical positions, it should not be hard to work out the effect a small change has. To look at an isolated position for 10-15 seconds is superficial.

Igor_Goldenberg
26-05-2006, 03:21 PM
Of course. However, if you consider that the team would have analysed many almost identical positions, it should not be hard to work out the effect a small change has. To look at an isolated position for 10-15 seconds is superficial.

Well, I beleive the team of Svidler, Bareev, etc. working for (and with) Kramnik had some experience in analysis.

Desmond
26-05-2006, 03:33 PM
Well, I beleive the team of Svidler, Bareev, etc. working for (and with) Kramnik had some experience in analysis.

Your point being...?

Igor_Goldenberg
26-05-2006, 03:56 PM
Your point being...?

If they could afford only 15 seconds per position, it means they have to go through a lot of variations!

Desmond
26-05-2006, 05:17 PM
If they could afford only 15 seconds per position, it means they have to go through a lot of variations!

Granted. It might have been better put that they spend (for example) 10 minutes of a particular variation than 10-15 seconds per position.