Author: Robert Hyatt
Date: 10:01:04 08/14/98
Go up one level in this thread
On August 14, 1998 at 07:53:51, Tom Kerrigan wrote: >Sounds logical, but I recall a graph in "Chess Skill in Man and Machine" (one of >the few interesting/useful things in the book...) that showed 1 ply = 200 rating >points, up through Deep Thought. > >Of course, using this rule of thumb to compare two programs in the same "era" is >fairly pointless, but I think that over long periods of time, the rule of thumb >may still hold true. > >-Tom > I think this might be "broken" today. IE in 1977 or so, when that was written, a ply cost us about a factor of 5.5, because there was no null-move search nor selective stuff other than a few search extensions. Today we use null-move R=2 to drive the branching down to 2-2.5, which means a ply today is really "less effort" than a ply 15 years ago. And, as a result, is probably < 200. I'd guess closer to 100 rating points per ply, perhaps, maybe less, since the most recent estimates show 70 elo for a doubling of speed, which is nearly another ply today... >On August 13, 1998 at 18:24:21, Robert Hyatt wrote: > >>On August 13, 1998 at 17:59:03, Jeff Anderson wrote: >> >>>I do not think that today a graph like that would be relevant. Some programs >>>sacrafice plies for a much more thorough evaluation. Other programs sacrafice a >>> thorough evaluation for speed. If you had a chart like this, a program that >>>searched slow would but evaluated well would do very well on a chart like that, >>>while a fast program that didn't have as complex evaluation would do poor. But >>>if you actually had a match between these two programs they two extremes would >>>probably even out. >>> >> >> >>two often-used measures, NPS and depth, are really only good for comparing >>A to A, ie the same program at a faster NPS or deeper depth. I believe that >>programs play better as they go faster, if the speed is gotten by either >>faster hardware or more efficient programming, rather than stripping something >>out or taking a "shortcut." >> >>I'd suspect that *everybody* agrees, because if you notice at every event where >>a computer plays another computer or a strong human, the computer operator makes >>every possible effort to get the fastest machine possible. >> >>But you can't compare NPS between two different programs and conclude anything >>about their "strength" based only on NPS. Ditto for "depth" because everyone >>has a different "meaning" for depth=10 plies, for example... >> >> >> >> >> >>> >>> >>>On August 13, 1998 at 06:12:37, Leon Stancliff wrote: >>> >>>> Can anyone tell me? >>>> >>>> A few years ago I saw a graph of ply level in the middle game versus >>>>anticipated rating. Has anyone attempted to do this same thing recently? We have >>>>estimates of the approximate rating increase per ply, and approximate rating >>>>increase with doubling of speed. >>>> >>>> Obviously selectivity, opening book and hashtables make a difference. But I >>>>feel certain someone has investigated the question I have proposed. What ply >>>>level was DeepBlue reaching in the Kasparov match? What ply level was Rebel 10 >>>>reaching in the longer games of the Anand match? >>>> >>>>Leon Stancliff
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.