Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: An Experiment that disproves Hyatt's 1000X NPS Theory

Author: Robert Hyatt

Date: 19:26:53 09/18/05

Go up one level in this thread


On September 18, 2005 at 16:14:10, Uri Blass wrote:

>On September 18, 2005 at 15:43:28, Chan Rasjid wrote:
>
>>On September 18, 2005 at 14:34:56, Uri Blass wrote:
>>
>>>On September 18, 2005 at 14:12:11, Chan Rasjid wrote:
>>>
>>>>This means there is NO-MATCH whatsoever with a 1 - 1000x nps difference.
>>>>Fruit's seems to sacrifice complicated (undue) evaluation in favour of high nps
>>>>and makes the top.
>>>
>>>I disagree
>>>
>>>Fruit has a very strong evaluation function.
>>
>>Firstly, I don't make definite statement but more of a guess, ie seems so to
>>me..The one think I have not examined yet in details is Fruit's evaluation.
>>
>>For your info, I think you made the same observations that I just found and
>>more of the secret of Fruit, ie data/structure organization, locality of
>>data,etc. eg. After I got the hint, when in domove(), I make(myself) copies of
>>board->sq[] for from, to, capture/ep with sqFrom, sqTo, sqCatureEP so that I
>>don't have cache jumping back to access board->sq[256]. Then I found Fruit does
>>the same. I believe the has an extra board->pos[256] for locality of data within
>>cache.
>>
>>
>>Strong evaluation may just mean good evaluation for:-
>>1) most basic standard stuffs that definitely do not harm, but gainful
>>2) b/r/ QUEEN mobility
>>3) attack king / king safety.
>>
>>Some others have full eval()for every pins. Does Fruit do any pins ?
>>If not then probably terms/factor wise, Fruit's evaluation is only professional
>>basic.
>>
>>>
>>>Most of the improvement from Fruit1.0 to Fruit2.1 is because of better
>>>evaluation.
>>
>>Of course eval() pays when optimal but after it is near optimal IT CANNOT
>>BE IMPROVED FURTHER. You can easily make it pays negative.
>>So we cannot be sure the reason Fritz, shredder lost is NOT because they wasted
>>critical time in pursuing complicated eval() that pays negative.
>>>
>>
>>>Crafty is faster than Fruit in nodes per second so nodes per second is not the
>>>secret of Fruit.
>>
>>By what factor , x 5 ?
>>
>>Ask Diepeeven. Crafty has a little evaluation (may be true comparatively ?)
>>and on bitboard which is not the way YET(?).
>>
>>I would bet generally if Crafty is just nps 3 x , it will outdo Fruit, provided
>>of course there are no major problems that the chess programming world is not
>>aware yet.
>
>The difference between fruit and Crafty is more than 3:1 speed advantage.
>


Where are you getting your info?  From games I have been playing against Fruit
on ICC, it's NPS on the same procesor is faster than mine...  Of course it is
impossible to verify hardware online, but I can verify NPS since the thing
kibitzes...




>>
>>>
>>>Uri
>>
>>I did this experiment, something like this:-
>>
>>   if (board->ply_nb <=	16)// dont eval if above ply 16
>>   eval_piece(board,mat_info,pawn_info,&opening,&endgame);// mobility is here
>>
>>   if (board->ply_nb <=	16)// dont eval if above ply 16
>>   eval_king(board,mat_info,&opening,&endgame);
>>
>>   //passers allowed
>>   eval_passer(board,pawn_info,&opening,&endgame);
>>
>>   if (board->ply_nb <=	16)// dont eval if above ply 16
>>   eval_pattern(board,&opening,&endgame);// very simple stuff
>>
>>Those are fruit's major eval().
>>It did not seem to weaken Fruit at all, wins ruffian,aristarch etc.
>>
>>Rasjid
>
>What time control did you use?
>At fast time control you do not get 16 plies in most lines.
>
>Uri



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.