Author: James T. Walker
Date: 13:14:34 07/21/05
Go up one level in this thread
Babelfish translation: In the third round we could be present at an historical fact: the computers also can be mistaken! Shredder 9,0 faced blacks MY Pablo Lafuente, and after 19 plays of a Nimzoindia Defense the game very was balanced, with probable tables of not mediating anything extraordinary. And the extraordinary thing happened. They already say the Laws to it of Morphy: "If something can fail, it will fail". Lafuente played 19.Axb7, changing the bishops of white squares, and Shredder, after meditating more than 3 minutes (to see graph of the registry of the game), answered with incredible the 19... Tfd8. If, was not about "mouse slip" or error of the operator, but a legal play made by the computer. While in a beginning Dxb7 analyzed obvious the 19..., in depth 20 (plys calculated 20 ahead, that is to say, 10 plays in brute force in all the variants), soon it began to analyze 19... Tfd8 with valuation of minimum advantage of the blacks, following with 20.Dc2 () Dxb7, etc. The surprise of Lafuente was capital, as much that it asked to me if that were the play that the computer had done. I did not have more remedy than to answer affirmatively, since one was a legal play and the rules do not allow the operator to modify what the computer in the monitor plays... Lafuente thought some minutes, and Shredder followed its analysis with valuation of balance, arriving until depth 23. After 20.Ac6, the valuation changed "to decisive advantage of the white abruptly". I decided to continue the game to see if one were some failure in the hardware, or if that error became to repeat. The game finalized normally in play 59, with victory of the young Argentine teacher. After finalizing the game, I returned to place the tactically important position in the board, in that same equipment, and then I already had left clear as it was the reason of the error: the tables hash. Let us see in question: All the modern programs of chess use the calls tables hash or tables of transposition to increase their force. In these tables analyzed positions are stored already. In the subsequent calculations the program verifies if the positions that appear are or in the table or no. If they already were, there is no necessity that it returns to study and to value the position again. This technique saves long time and the program can obtain with him greater depth of calculation. Shredder returned to consider, instantaneously, 19... Tfd8 like the best answer, arriving until depth 22... Evidently, the program read the valuation stored in the table hash, and this valuation or calculation was mistaken. Failure of some routine of the program or some disadvantage with "the physical" memory (ram) of the computer? What I create is that the possibility that that failure is repeated must be something asi like 1 in 1 million! The equipment worked (and it continued working normally) and the program never "was hung", but the certain thing is that escribió/leyó an incorrect value in the table hash, and that was the cause of the terrible error.
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.