Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Matevalues, hashtables and iterative deepening

Author: Odd Gunnar Malin

Date: 05:48:20 11/30/01

Go up one level in this thread


On November 30, 2001 at 07:25:42, Erik Robertsson wrote:

>I´ve been trying to correct some bugs wich occur when storing mate values in the
>hashtable.
>
>If the position is evaluated as a mate value at the end of an iteration in the
>iterative deepening, I stop searching and execute the best move. In some
>positions however, I find that due to finding positions in the hashtable it can
>result in the engine alternating between two different mate variations ending in
>a quite embarrassing draw by repetiton.
>
>I've tried to solve this by not automatically stop the iterative deepening
>unless I've found a shorter mate than found in the previous move. Is there a
>better way of solving this? I remember reading something about this some year
>ago, but cannot find it.
>
>I use the draft = *infinity* for mates wich are not bounds which should result
>in that most positions with mate should be in the ht in the next search. Maybe
>it has something to do with the bounds, but I´ve tried Crafty's MATE-300 bound
>and it sometimes result in getting the evaluation of the position as MATE-300.
>
>This has kept me awake a few nights, please help or I will go insane.


I guess you allready do this when saving hash:

if (bestscore>(MATE-MAX_PLY))
  hashscore=bestscore+ply;
else if (bestscore<(-MATE+MAX_PLY))
  hashscore=bestscore-ply;
else
  hashscore=bestscore;

And the opposite when you reading hash:

if (hashscore>(MATE-MAX_PLY))
  bestscore=hashscore-ply;
else if (hashscore<(-MATE+MAX_PLY))
  bestscore=hashscore+ply;
else
  bestscore=hashscore;

If you use fail-soft alpha-beta it could be the same error that I had in my
little engine.
I was returning alpha instead of bestscore in  one of my cutoff.

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.