Author: Robert Hyatt
Date: 07:39:42 05/15/99
Go up one level in this thread
On May 15, 1999 at 03:59:34, Mark Young wrote: >On May 14, 1999 at 16:58:03, Robert Hyatt wrote: > >>On May 14, 1999 at 11:07:36, José Carlos wrote: >> >>>On May 14, 1999 at 05:24:20, Peter Hegger wrote: >>> >>>>Hello >>>>Let's say that today's best programs, Fritz, CM6000, junior etc.. are playing at >>>>the 2450 level at 40/2 when they've got hardware capable of knocking off .5M >>>>nps. I don't think this is too outlandish an assumption. >>>>If you double this speed 8 times over you arrive at 128M nps. This is in the >>>>same ballpark as this new proposed screamer of Hsu's which it is estimated will >>>>knock off 120M nps on a multi-processor platform. >>>>I've seen in other threads that doubling speed will increase performance >>>>anywhere from 30-70 points per doubling. For argument's sake and to split the >>>>difference I'll assume that 50 is likely pretty close. Using 2450 as the base >>>>this would translate into an elo of 2850 give or take a bit. >>>>Is it really possible that a machine which is stronger (marginally) rating wise >>>>than the world champion is right around the corner. Or am I missing something >>>>here in making this estimate? >>>>In any event I'd love to see Kasparov tackle this baby in a 40/2 24 game match. >>>>Bets anyone? :) >>>>Regards >>>>Peter >>> >>> >>> The increment of peroformance doubling speed is more little as speed >>>increases. Doubling speed allows, usually, to go one ply deeper. So it's very >>>different to go from ply 7 to ply 8 than to go from ply 50 to ply 51, isn't it? >>> >>> José C. >> >> >>You need to read the ICCA Journal. There is lots of evidence (now) that >>going deeper does indeed lead to better play.7 to 8 is clearly going to do >>more than going from 50 to 51. But 7 to 8 might not be any better than >>going from 14 to 15 or even 19 to 20, based on experiments both I and Ernst >>did. Programs _still_ find better moves at deeper depths, even when the >>depth is increased from 14 to 15 or 15 to 16. > > >Why must there be experimental evidence to show that programs that are "non-null >movers" and have a progressive full width search component get better as their >search gets deeper? We know that chess is finite and 100% tactical and thus in >general the question of ply depth can be answered through mathmatics and no >experimental evidence is needed. However, for individual programs, experimental >data may indeed hold value. > >In simple terms, the deeper a program searches the less chance for a horizon >effect to bite you thus breaking the balance of the game turning a draw or a won >position into a losing one. > >Now, actully calculating an accurate average for how much an additional ply >added to any given search depth will boost your rating would be a tremendous >undertaking experimentally, since there are so many variables to be taken into >account. But, experimentally is the only way to proceede since we've currently >been unable to solve the little problem of chess mathmatically as of yet. I hope I didn't imply that. I was only pointing out that there is now quite a 'pile' of experimental data that suggests that every additional ply improves a chess program, and that there is little evidence to support the concept of 'diminishing returns' or the oft-used 'tactical sufficiency' concept. IE in the "Crafty Goes Deep" and "DarkThought Goes Deep" papers in the JICCA, we found that crafty finds a better move going from 13-14 plies about as frequently as it finds a better move going from 9-10 plies. Dark Thought produced similar data...
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.