Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Is position learning the key in shredder 7.04

Author: Torstein Hall

Date: 04:19:43 05/29/03

Go up one level in this thread


On May 29, 2003 at 02:55:04, Rupesh wrote:

>Shredder 7.04 shows a curious behaviour. While playing on playchess, i have
>observed that shredder's eval has a tendencey to change signs dramatically (
>like -2.4 from +1.3, many examples like this abound). While it is true that the
>program does not show this kind of behaviour if the same position recurs again
>and it does try a different move (perhaps using its learn file).
>
>So the program plays really badly initially and becomes better gradually after
>playing a reasonably large number of games. I have a plr ( the learn file of
>shredder 7.04) of 4.9 MB which is considered to be very big. Still shredder has
>not stopped the habit of changine the sign of its eval.
>
>What i really suspect is that shredder's evaluation function is not robust and
>often misses completely to evaluate a chess position at crucial time.I bneleive
>this happens with shredder significantly more often than other programs. But
>over repeated games against a fixed number of opponents ( it learns to expect
>the kind of positions that might arise and becomes better)..but if it faces a
>new opponent ( may be the same program on a different hardware and setting), it
>again has to learn before it starts beating the opponent program.)
>
>It has been a curious fact that shredder starts doing well after playing many
>games against a fixed opponent. I t almost always loses initial games against
>any program ( mean inng any good program like f8, df7).

I do not know Shredder very well, but I find it not very probable that the learn
file should should affect play a lot. As there is quite a number of positions in
chess and you would have to play a laaaarge amount of games to hit many learned
positions during a game. But perhaps SMK has found a new way to generalise
position learning, so the position only had to be similar and not 100% equal,
but I doubt it.

Torstein
>
>To my mind, it shows that shredder depends on learn file heavily. Which is
>something which dissappoints me as i look for a program which has a good
>evaluation function rather than learn function.
>
>Any remarks will be very muchg appreciated.
>
>Amitesh



This page took 0.01 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.