Author: Christophe Theron
Date: 10:40:09 09/28/00
Go up one level in this thread
On September 28, 2000 at 08:09:22, leonid wrote:
>On September 27, 2000 at 17:05:06, Christophe Theron wrote:
>
>>On September 27, 2000 at 06:34:50, Aaron Tay wrote:
>>
>>>On September 27, 2000 at 04:07:45, Uri Blass wrote:
>>>
>>>>On September 27, 2000 at 03:20:21, Aaron Tay wrote:
>>>>
>>>>>On September 26, 2000 at 15:01:19, Christophe Theron wrote:
>>>>>
>>>>>>On September 26, 2000 at 09:31:30, Bas Hamstra wrote:
>>>>>>
>>>>>>>It seems the current LG versions are nps wise down to slightly faster than
>>>>>>>Crafty :) I play it all the time at FICS. It seems is evolving to a slower and
>>>>>>>more knowledgable program. As is mine, BTW :)
>>>>>>>
>>>>>>>On September 25, 2000 at 01:38:15, Jouni Uski wrote:
>>>>>>>
>>>>>>>>In my moderate AMD K6-2 450Mhz and 50 MB hash it exceeds easily 1000knps in
>>>>>>>>tactical positions and sometimes goes over 1300knps. Of course I know this means
>>>>>>>>almost nothing to playing strength, but still it's unbelievable. I wonder can
>>>>>>>>it be true nps value or has Michael B. his own node definition...
>>>>>>>>
>>>>>>>>Jouni
>>>>>>
>>>>>>
>>>>>>
>>>>>>Don't mix up "slower NPS" and "more knowledge". They have nothing to do with
>>>>>>each other.
>>>>>>
>>>>>>In my example, I have said how I could make Tiger much faster just by turning
>>>>>>off some obvious selection schemes. If I turn these selection algorithms back
>>>>>>on, then the program is suddenly much slower. It is also much stronger.
>>>>>>
>>>>>>But the amount of "knowledge", as most people understand by "knowledge" has NOT
>>>>>>changed at all. I mean that the evaluation of chess positions is still exactly
>>>>>>the same. If the program does not understand a knight outpost in the fastest
>>>>>>version, then it still does not understand it with the slower version.
>>>>>>
>>>>>>When a new version of a known program is released, some people will look at the
>>>>>>NPS and say "Oh, the NPS of the new version is lower, so I guess more knowledge
>>>>>>has been added to the program". Bullshit.
>>>>>>
>>>>>>
>>>>>>
>>>>>> Christophe
>>>>>
>>>>>
>>>>>Sorry but talking as a person who knows nothing sbout chess programming, isn't
>>>>>it true that the more complicated evalution functions you add, the slower the
>>>>>search becomes to reach the same depth?
>>>>
>>>>The main point is that there are other reasons to be slower(more complicated
>>>>search rules or different definition of nodes) and you cannot know the reason
>>>>for being slower so you cannot say that a program that needs more time to >search the same number of nodes has more complicated evaluation function.
>>>
>>>>The discussion was about nodes per second and not about being faster or slower
>>>>in reaching the same depth.
>>>>
>>>>You can be slower in reaching the same depth by not prunning lines by null >moves or other prunning ideas so the depth tells nothing about the complexity >of evaluation.
>>>
>>>Oops poor choice of words..But generally more complicated evalutation/knowledge
>>>= lower NPS?
>>>
>>>
>>>Of course as the thread above says, you can have lower NPS for other reasons..
>>>
>>>
>>>Christophe mentions "switching off obvious selection schemes" to speed up Tiger,
>>>what exactly does that mean?
>>>
>>>For example,
>>>Does he mean switching from a more complicated adaptive null move pruning method
>>>where R=2 or 3 depending on sitution to a simpler say null move R=2 througout
>>>will increase NPS also?
>>>
>>>
>>>But does it really cost so much more?
>>
>>
>>
>>No, I wasn't talking of playing with null move. Null move is only ONE
>>selectivity algorithm, but there are DOZENS of other ways to be selective and
>>gain strength from selectivity.
>>
>>
>>
>>
>>>So with all things equal , as your search becomes more complicated, NPS
>>>drops..Eg from a simple search add Killer moves, Quience search then null move
>>>and NPS drops furthur??
>>
>>
>>
>>Yes. Adding a QSearch to an existing program drops the NPS tremendoulsy for
>>example. Because generating only capture moves takes more time in average per
>>generated move than generating all the possible moves in a position.
>>
>>You add QSearch => your NPS drops. But the program plays much better.
>>
>>This is just an example, but there are a lot of others.
>>
>>That's why I say there is no relationship between NPS and playing strength, and
>>no relation between NPS and amount of knowledge.
>>
>>As far as I know, null move has almost no impact on the NPS.
>>
>>
>>
>> Christophe
>>
>>
>>
>>
>>>>Another point is that it is possible that a more complicated evaluation will
>>>>help you to get bigger depth because the evaluation may help you to get a >better order of moves.
>>>
>>>Agreed .
>>>
>>>NPS may be irrelevant. Kasparov has a low NPS compared to computers but because
>>>of better "move order" so beta cutoffs come quickly and singular extensions , he
>>>can usually search a deeper depth..
>>>
>>>>Uri
>
>So far as I could see in my program, NPS diminution is due mainly to the less
>efficiency in move generator, when only part of its moves are used. When all its
>moves are used (minimax), then all the data found for verification of legality
>of those moves is used 100%. When number of moves used drops from 100% to the
>20% or 15%, NPS goes down to around 3 or 4 times. And those numbers do not take
>in consideration "additional knowledge" that will drop NPS even farther. Only
>reason for last drop is different.
>
>Leonid.
Yes, I see you understand my point.
Christophe
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.