Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: learning

Author: Odd Gunnar Malin

Date: 15:19:03 04/08/05

Go up one level in this thread


On April 08, 2005 at 10:30:37, pavel wrote:

>On April 08, 2005 at 06:10:05, Odd Gunnar Malin wrote:
>
>>On April 08, 2005 at 04:38:07, jim r uselton wrote:
>>
>>>Novice question---do some computers have a learning function? Which computers do
>>>this and how do they do this? How do they learn?
>>
>>I assume you thinking of chessplaying program/computer here and how they learn.
>>
>>There are two types of learning, book learning and positional learning:
>>
>>Book learning.
>>If a chess program comes out of the book with a bad score and this stays bad for
>>a certain number of moves the program search in the book backwards to mark the
>>move lead to this as bad, (internal its easier just mark the end position in the
>>book as bad and use some minmax search to get the score at a branching point).
>>It also do the same when the game end to keep a score of if this line lose, wins
>>or draws to try to repeat won games.
>>This learning feature isn't too important for other than in computer
>>tournaments/matches. If you play as human and see that the computer do a bad
>>move in the opening you just correct the score for this move in its book so you
>>aren't bothered with it anymore.
>>
>>Positional learning
>>After each time the computer have finish a search an is ready to move it save
>>the move, position, searchdepth and score into a file for later use. Then before
>>each game it read this file into memory (usually just into the hash table) and
>>can take advantage of it in the search the next time the same position is on the
>>board. This should hopefully result in that it will select a slightly differrent
>>path if you repeat the same opening or playing some sort of thematic game.
>>This learning feature is a lot more important if you use your computer as a
>>sparring partner in your training session.
>>
>>There is of course other ways to do these two learnings, but it mayby it helps
>>you to understand when and where you should use them in your program.
>>
>>Odd Gunnar
>
>I think it would be a interesting option if a program could "learn" from
>analyzing games. I could let a program analyze a game for a certain amount of
>time per move and it would save the analysis (score, PV; just like position
>learning).

Engines that do positional learning in analysis mode like Shredder and Gandalf
can do this too.

>I don't know if it would help it's game, but I think it would be a nice option
>to have.

I can't neither see that this would help anything, the chance to get into the
same position is too small. So maybe if you play a match against the same
opponent and then analyze the games and do a new match.

There have been some try to adjust the evaluation function based on automatic
analyses (didn't Deep Blue do this).

Odd Gunnar



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.