Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: "Don't trust draw score" <=Is it true?

Author: Bruce Moreland

Date: 10:19:36 08/10/01

Go up one level in this thread


On August 10, 2001 at 06:34:29, Martin Giepmans wrote:

>On August 09, 2001 at 12:38:58, Bruce Moreland wrote:
>
>>On August 09, 2001 at 06:12:52, Martin Giepmans wrote:
>>best move is, say, Nc3. Score +0.10.
>>>
>>>I think this proves your point: drawscores influence other (nonzero) scores.
>>>However, along path B white has obviously not played the best moves.
>>>If white gets 0.10 instead of 0.40 there must be something wrong with either 2.
>>>Nf3 or 1. d4 or both.
>>>So, my question is: does this influence of drawscores minimax downto the root of
>>>the tree? My first thougth was no, my last (thought 31) is: I really don't know
>>>...
>>
>>The answer is yes, as anyone who has seen the fail-high/fail-low effect in a
>>vanilla program with nothing special except hash tables can attest.
>>
>>This was the subject of one of my first r.g.c.c. posts, in 1994.
>>
>>bruce
>
>
>What is this? Are you getting irritated? Maybe I'm stubborn but to me it's not

No more than usual.  It's like 6 days until plane trip, and that won't be fun.

>about who is right or wrong. I just want to be sure. And what anyone sees, well,
>500 years ago anyone could see that the earth was flat ...
>What happens in a search is obviously very complicated. Alfabeta, hashing,
>pruning, etc, all these things interact. We try to find our way by a kind of
>fuzzy reasoning, mixed with questionable induction from experience.
>It's like in the early days of physics or statistics. People often got wrong
>results because they had no exact mathematics or controlled experimenting to
>rely on. In this situation we have to be very careful. Any conclusion might be
>wrong. Do you agree?
>
>About the weird thing in your vanilla: I've also seen it, but that was
>*before* I implemented detection of repetition. It disappeared when I disabled
>pruning ... You see, there is the doubt again ...

My original version used no pruning of any sort, other than the safe alpha-beta
kind.  I got this problem when I added repetition testing.

Any sort of pruning based upon alpha or beta, and that includes null-move
pruning and all that, will lead to this problem as well.

Hashing will also cause it if:

1) You get a collision.
2) You cut off based upon the results of deeper searches.

All the good stuff makes your search unstable.

bruce

>Anyway, I understand I will see you in Maastricht, I wish you good luck but of
>course I hope I will mate you :)
>
>Martin



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.