Author: Russell Reagan
Date: 18:52:29 07/11/03
Go up one level in this thread
On July 11, 2003 at 21:27:58, Bob Durrett wrote: >Suppose the programmer were to deliberately load the hash table with false data >and then start the engine analyzing a given position. Could the false >information in the hash table cause the engine to produce the wrong move? Sure. Some tests have shown that invalid entries in the hash table have a very low chance of causing an error or a bad move to be played, but it is possible. >Corollary: Suppose the engine were given 24 hours [or more] to find the correct >move, after the programmer had loaded bogus information into the hash table. >Could it happen that the engine might NEVER find the correct move as a direct >result of incorrect information having been loaded into the hash table? I suppose so. Imagine an engine that used a scheme of "only replace if the depth of the new entry is greater than the current entry". If the current depth was 1 million, then nothing will ever be replaced. If every entry in the table had an inverted score for the resulting position of every legal move in the current position, then the engine would get all cut offs and play the worst move in the position, instantly. All of this is dependant upon a lot of things though, like all of the details as to how the programmer implemented various things. >I would very much appreciate an answer to this. I am experimenting with >different ways to use the engines in analysis and would like to know whether or >not I am doing anything right. It would really help to understand the details >of how modern engines use the hash tables. [Maybe different ways for different >engines?] I don't think there is much that you as the user can do to alter the hash table settings much to affect the moves it chooses (other than increasing or decreasing the table size). All other problems that you mention above would be on the programmer's end.
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.