Author: Mike S.
Date: 12:21:29 05/27/03
Go up one level in this thread
On May 27, 2003 at 14:20:25, Michael P. Nance Sr. wrote: >Correct Me if I'm wrong. Fritz learns,(It changes the wieghts of choosing an >Opening Book move ,from reading a Database),when You enble the "Learn form >Database"function of the Program. No, "Learn from Database" is a book editing function, not a book learning option. It's not what you mean. There are two methods of automatic permanent learning (mainly): 1. book learning, changes book move weights as you describe. AFAIK, in Fritz *only* the GUI handles this. IOW, not the engine edits the book file (but of course it's based on the engine's evaluations and results). 2. position learning, where the engine learns evals of specific positions under certain circumstances (i.e., when the original evaluation later turns out to have been very wrong :-). By that, an engine can (try to) avoid mistakes, but also can remember strong moves. This is handled by the engines themselves, and I guess in different ways. For reasons I don't quite understand, many people seem to think only of the book learning always, when they talk learning, and seem to forget position learning. Although, pos.learning is present nearly everywhere. For example, in Crafty and many other freeware engines. Ever read docs of such engines? Most often you'll have to configure, position learning on or off, etc... Also I wonder if many people never take a look which files appear in their ../Engines/.. folder. Owners of Hiarcs, Nimzo, Shredder will see files appear there *other* than ENG, DLL or INI. Files which weren't there, before the engine was run for the first time. These are the position learning files, where that learnt data is stored. Often, the extension is .LRN. From the engines I know, the following use position learing: Hiarcs 7.32 (8 too I guess), Nimzo 8 (7.32 most probably too), and Shredder. Many freeware engines can use pos.learning, i.e. WB-Crafty (not cb-native Crafty), Yace, Pharaon, Li.Goliath 1.5 ... but i.e. not Ruffian 1.0.1 nor List 5.04. Junior 5, CB-Tiger 14 or Fritz 7 didn't use pos.learning (yet); I cannot comment on the new versions. Position learning virtually guarantees, that if you play *from the same opening database position* (or from the same book variant) under the *same conditions* on the *same computer*, that two engines each will nearly always play a different game than previously. There may be other reasons for this (sometimes), but obviously the main factor is the position learning. It can also be the answer to the "secret", when solving times of test positions can't be reproduced... because the second time, it solves so much quicker?! :-) It's interesting to notice, that sometimes the winner will deviate first (and eventually loose then :-). - But in general, position learning is of course a very good way to try to avoid mistakes, using information which occured later in the game than when the mistake happened. Recently, I played a 60 games blitz match between Hiarcs 7.32 and Nimzo 8, not using an opening book but *only one specific opening variant* for each game. Each game was different! Guess why :-) Regards, 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.