Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Ply Depth in relation to Elo again...

Author: Robert Hyatt

Date: 06:05:15 05/23/00

Go up one level in this thread


On May 23, 2000 at 01:19:27, Mark Young wrote:

>On May 22, 2000 at 23:02:36, Robert Hyatt wrote:
>
>>On May 22, 2000 at 22:24:20, Mark Young wrote:
>>
>>>On May 22, 2000 at 18:35:44, Joshua Lee wrote:
>>>
>>>>This is the problem i get juggling those numbers....
>>>>Deeper Blue                1997   (15at a minimum)17-18 Ply  2817 (USCF)
>>>> 200 MNPS 2.048Thz
>>>>
>>>>2817USCF -50 or -100   2767 or 2717FIDE  (not to get sidetracked but keep this
>>>>in mind -107  2660 or 2610)
>>>>
>>>>You said DB2 is +1 ply because it is 2x Faster let's keep in mind the minimum
>>>>Ply as well
>>>> Deep Blue                    1996     (14Ply Min 16-17 Ply (*)14 Ply
>>>>2642(FIDE?)
>>>> 100 MNps 1.024Thz
>>>>
>>>>512Ghz would be 50MNps  and (13Ply Minimum)15 Ply so 2642-50= 2592   or -60
>>>>=2582
>>>>or is it -90 =2552???
>>>>
>>>>i could really confuse things by considering Deep Blue's Positional Strength
>>>>as  plain and simple a Programs strength is just it's positional play (which
>>>>isn't a fact ) people play program X but only positionally. They keep the
>>>>position devoid of tactics...
>>>>so let's say 2535- 50 =2485  or -60 = 2475   or is it 2445?
>>>>
>>>>Something tells me that this is going to end up Showing that PC's are around
>>>>2000 which isn't the case!!
>>>>
>>>>what i was saying above is that i don't know that we should say programs are
>>>>-107 or so because lack of positional strength. I got this number by not
>>>>counting the first game of the DB1 match  Kasparov's rating at the time was 2775
>>>>which is in itself way inflated from the 70's but that is another topic.
>>>>I remember someone saying how if you disregarded the first game the results show
>>>>that even at this depth the positional strength of DB was at 2535 (-107)
>>>>
>>>>You said : doubling the cpu speed is generally said to produce 50-60 rating
>>>>points.Since the typical effective branching factor is around 3.0, every time
>>>>the
>>>>speed is tripled, we get another ply, and using the 60 point figure above,
>>>>a ply would be about 90 rating points, roughly.  But there is nothing that
>>>>says that as we go deeper this doesn't taper off.  Nor is there anything that
>>>>says that as we go deeper, the gain doesn't actually get larger...
>>>
>>>You are wrong, there is a lot of "say" that shows as you increase ply depth the
>>>next ply you increase gets fewer and fewer rating points.
>>>
>>
>>
>>Where?  IE I have seen this _said_ several times.  But I have never seen it
>>_shown_.  Two good examples to the contrary are the "crafty goes deep" and
>>"dark thought goes deep" published in the JICCA.  Both suggest that at least
>>until we reach depth 15/16, each additional ply is revealing new and more
>>important information about a position...  and resulting in steadily improving
>>play...
>>
>>I think the programs are getting stronger every year, even if nothing is changed
>>but faster hardware...  and it is reasonably linear...  so far.
>
>It is....Hmmm What is the rating jump going from say 3 plys to 4? What is the
>rating jump going from 7 plys to 8? What is the rating jump going from 11 plys
>to 12 plys. This we have and can get data on, and you know as I do it is not
>near linear. As the ply depth get deeper you gain less by adding one more ply.
>



I don't know that at all.  I know of two experiments that took two well-known
programs and tried them on a set of game positions (over 300).  And then the
difference from ply-to-ply was measured.  And at depth=15 the programs were
changing to better moves just as often as they did at depth=3.

Which is why many don't think there is any 'flattening' of the rating curve
at all.  I'm not sure that there is no "diminishing returns".  But I am sure
that there is some evidence to suggest there is none.  Both papers were
published in the JICCA.



>Now it is true as you go deeper the curve does flaten out and it because more
>linear, but the trend *so far* is always for less ratings gain for the next ply
>searched.
>


Based on what?  The original tests were always done with program A playing
program A at 1 more ply.  I am not convinced that means anything at all.  The
more recent tests were program A vs many positions taken from a game played
by a human.


>At todays point on the curve if a ply is worth 90 rating points, do you really
>think for example going for 25 to 26 plys will get you anywhere near a 90 rating
>point gain with all things being equal except ply depth.
>
>I think not.


First, I don't think "90" is a "magic number".  Bruce found that Ferret
searching very shallowly could hold a rating > 2200 on ICC.  At the time,
his program was rated over 2800 in a real rating there.  Which means each
ply was worth much less than 90.    So "N" points per ply, linearly?  I think
it is definitely possible.  And yes, it is possible that some programs begin
to 'flatten out' while others don't.  I don't see anything that suggests that
can't happen either.



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.