Author: Mike S.
Date: 12:36:01 05/06/03
Go up one level in this thread
On May 06, 2003 at 08:32:35, emerson tan wrote: >Does a chess computer really learn from its mistake and grasp the idea of its >mistake to avoid it the next time or it just make a statistics of which opening >variation it loses and avoid them the next time? Yes, both. There is (1.) Book learning, which changes the probabilities of opening moves (not immediatly avoid lost openings, but change how often book moves are chosen). (2.) Position learning, which saves positions where big evaluation changes occured later in the game (i.e. mistakes = moves which were evaluated "wrong" at first, when played in the game). Not every top engine has position learning, i.e. Fritz, King, or AFAIK Tiger and Junior still don't have it, but Shredder, Hiarcs and Nimzo have it, many freeware engines too (Yace etc., but i.e. not Ruffian). So, to be competitive book learning is sufficient obviously - when lost openings are avoided, positions from later in the lost game won't occur again anyway... (except some transpositions then and when). *BUT* these engines have no chance to alternate their play starting from the same variant, when fixed opening variants are used (opening databases), like the Nunn or Noomen opening sets. This isn't much of a problem though, when engines don't play the *same opponent from the same opening more than once*, at the same conditions. But theoretically, if that is done with 2 engines which don't have position learning, it must end in a double of the first game... It's astonishing how effective the *position learning* can be: I've just run some long matches - up to 60 games! - with 2 engines each, from the *same opening variant* each, same conditions for each game (alternate colors). You'd think this will give many doubles. But it turned out that every game was different! Most often, the alteration came relatively early, not just somewhere at the end of the game. It also happened often, that an engine didn't repeat a continuation it had previously won with - and then lost... (but I don't doubt that if it would have tried to repeat the win, the other engine would have played an alternative sooner or later). This is an interesting way to check opening novelties btw. :o)) Ragards, M.Scheidl
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.