Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: interesting idea (some food for thought)

Author: Robert Hyatt

Date: 11:08:25 09/09/02

Go up one level in this thread


On September 09, 2002 at 13:30:59, Miguel A. Ballicora wrote:

>On September 09, 2002 at 03:13:09, José Carlos wrote:
>
>>On September 08, 2002 at 22:24:06, Miguel A. Ballicora wrote:
>>
>>>On September 08, 2002 at 21:17:47, José Carlos wrote:
>>>
>>>>On September 08, 2002 at 20:46:44, Miguel A. Ballicora wrote:
>>>>
>>>>>On September 07, 2002 at 11:13:20, Robert Hyatt wrote:
>>>>>
>>>>>>
>>>>>>Jose made a really good point about observed data vs measured data.  After
>>>>>>thinking about it for a bit, I decided that it is a point strong enough to
>>>>>>change the way we think about "measured" and "observed".
>>>>>>
>>>>>>Some examples:
>>>>>>
>>>>>>speed.  Impossible to measure.
>>>>>>
>>>>>>For example, your automobile (newer vehicles) compute speed by counting the
>>>>>>revolutions of the tailshaft (output) of the transmission, then factoring in
>>>>>>the rear-end ratio and the circumference of the rear wheels.  It _computes_
>>>>>>the speed from that.
>>>>>>
>>>>>>A radar measures the frequency change in a radio signal as it bounces off
>>>>>>a moving target and _computes_ the speed based on the frequency change.
>>>>>>
>>>>>>A GPS observes to "positions" in terms of lattitude and longitude, uses some
>>>>>>geometry to compute the distance between them, and uses a clock to measure the
>>>>>>time to cover that distance, and displays speed.
>>>>>>
>>>>>>So Speed can't be measured directly, it has to be computed.  And this isn't a
>>>>>>surprise since speed is defined as distance over time.
>>>>>>
>>>>>>Brightness.  (of a light, not a person.  :)  )
>>>>>>
>>>>>>This is a direct measure of an electrical signal produced by some sort of
>>>>>>device (photo-resistor, photo-cell, optical transistor, etc) and then that
>>>>>>voltage is used to compute a brightness level in Lumens...
>>>>>>
>>>>>>Loudness (sound).
>>>>>>
>>>>>>Ditto.
>>>>>>
>>>>>>NPS.
>>>>>>
>>>>>>nodes searched divided by time in seconds.  Computed.
>>>>>>
>>>>>>Speedup
>>>>>>
>>>>>>one-processor time divided by the N-processor time.  Computed
>>>>>>
>>>>>>We really don't have a lot of "observed" data nowadays.   Some, yes.  Where
>>>>>>were you at 8pm last night.  But more is computed...
>>>>>>
>>>>>>Which means if we start to define observed vs computed, we don't end up with
>>>>>>very much in the "observed" column.
>>>>>>
>>>>>>In a chess program I can count nodes and "compute" time (end-time minus
>>>>>>start-time) and then compute a nps value.  I can measure run-time and compute
>>>>>>speed-up.  But I can't directly measure speed at all.
>>>>>
>>>>>Not really, NPS is a direct measure. You do not measure nodes, you "report"
>>>>>them. In other words, there is no error in the measure of nodes.
>>>>
>>>>  Not quite correct, IMO. Say you start iteration 11. You generate moves at the
>>>>root, then pick the first one. Go along the PV and go back in the tree. Let's
>>>>say you're in ply 6. So plies 1-5 are PV moves and from 6 on you have searched a
>>>>subtree. You test time and find you ran out, so you decide to stop searching.
>>>>Did you completely analyze the root node? I don't think so. You analyzed it
>>>>partially. Same for node in the PV until ply 5.
>>>>  In parallel search there's something more. When you decide to stop the search
>>>>some processors might be generating moves, others might be evaluating, and so
>>>>on.
>>>>  So "nodes searched" are not a discrete meause, thus they must be measured, not
>>>>just reported.
>>>
>>>They (nodes) might not mean what you really wanted to mean, but you define them
>>>and count them. For instance, number of times search is called or number of
>>>times makemove is called, whatever. You measure how long does it take to do a
>>>certain number of those events. Once those events are counted, they have no
>>>statistical error. If they do not represent what you want is a total different
>>>issue.
>>>
>>>Regards,
>>>Miguel
>>
>>  Of course. I can define anything to make my life easier, but I guess that's
>>not the point. You can define speed as a constant (distance from Madrid to
>>Buenos Aires) divided by the time I need to do the travel, but that's not the
>>definition of speed, just a particular case.
>
>I was talking about the definition of nodes. You choose the definition, once you
>do it, you "count" them.
>
>>  Also in a chess game, you don't do a search until some node count (whatever
>>the definition is) is hit and then stop. Your search decides _when_ to stop
>
>Yes I do :-) every 256 nodes I check the time. I never check what the time is
>unless I am in (nodes%256)==0. Except when a move was found, nodes are checked
>(reported) and then time is checked. I never check the time and then measure the
>nodes. I thought most of the people did it similarly.



The problem is _you_ check the time every 256 nodes.  Doesn't mean the O/S
_updates_ the time that often.  In fact, on PC machines it certainly does
not.  And about the best accuracy you get today is 10 milliseconds +/-...
And that not on a PC

>
>>based on time. So your node count is not constant. We could argue whether it is
>>a discrete measure or not, but certainly it's not constant.
>
>It is a discrete measure and because of that, no matter how you measure the
>nodes, the error is +/- 0 (unless there is a bug in the counter :-). In error
>propagation is treated as a constant, that was my point.
>
>Regards,
>Miguel

If there is no error then you can reproduce it, correct?

Then we have _another_ problem with SMP nodes...


>
>
>>  José C.



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.