Author: Uri Blass
Date: 15:02:31 07/17/02
Go up one level in this thread
On July 17, 2002 at 16:38:53, Sune Fischer wrote: >On July 17, 2002 at 15:55:12, Dann Corbit wrote: > >>On July 17, 2002 at 15:25:34, Sune Fischer wrote: >> >>>On July 17, 2002 at 14:32:11, Dann Corbit wrote: >>> >>>>On July 17, 2002 at 14:13:57, Sune Fischer wrote: >>>>[snip] >>>>>I would love to see a G/5 rating list compared to a G/120 list, I think most >>>>>good engines does about equally well in all time controls. >>>> >>>>This is a common contention. I have heard it said by such experts as Christophe >>>>Theron and Gian-Carlo Pascutto. However, it is clearly and transparently wrong. >>>> >>>>Assertion: >>>>No two chess programs with non-identical search and evaluation have *exactly* >>>>the same branching factor. >>>> >>>>If this assertion is true, then it is transparently false that chess programs >>>>play with the same level of excellence at all time controls. >>>> >>>>Or, stated another way, for every chess program, at *some* point, the branching >>>>factor of that program will totally dominate the play of the program. >>>> >>>>Draw a curve of: >>>> >>>>C0 * exp( C1 * x) >>>>verses >>>>C2 * exp( C3 * x) >>>> >>>>and as long as constants C1 and c3 are not identical, at some point the equation >>>>with the smaller constant in the exponent will dominate. If (by chance) C0 or >>>>C2 is also smaller for the smaller exponent modification, it is possible that >>>>one program will always be better everywhere [within reason]. >>>> >>>>However, the assertion that the entire family of curves never have intersections >>>>is [with no doubt whatsoever] utterly false. >>> >>>Yes I know the math (I've posted it myself:). >>> >>>The point is, can you give examples where program X is better at blitz and >>>program Y better at standard time (on ~1GHz)? >>>I rarely see these lines cross, and if they do it is probably only between very >>>equal opponents or weak engines that have bad tuning. >>>IIRC people on this board have refuted the notion, that Hiarcs should be a bad >>>blitzer just because it is a slow searcher, for instance. >>> >>>I prefer to play around 50-100 games to see if the changes were for the better >>>or worse, this would take weeks if I had to play at standard time. >>>I doubt there is any significant difference (other than a lot of wasted time). >> >>The superfast games I posted (games in 15-60 seconds) are obviously pure crap. >> >>At some point, that will stop being true. Exactly where is the transition? > >No it will not actually. >They search deeper, that's all, there is no transition. >They go from utter crap/random moves to perfect play, I don't think the >difference between 10 plies and 13 plies is very significant in that respect :) > >10 years ago G/120 was fine on the computers back then, today that nodecount is >crap and G/120 is fine, 10 years from now todays nodecount will be crap and we >will expect to see 30 billion nodes analyzed per move, or else we think of it as >crap :) > >>I think the linear scaling of chess engines is not something you can rely on. > >I don't know if it changes their relative strength if I let them search longer, >I have a feeling it doesn't matter a whole lot (that is why I would love to see >the two lists). It is possible that it is not important for the best programs but I can show you examples that it important for weaker programs and your program is still not in the team of the best programs. I think that movei is clearly better than monarch if you use 40 minutes/40 moves time control(see the results of leo's tournament) but it is not the case if you use faster time control. Monarch and faile are at similiar level at 40/40 based on Leo's tournament. It is not the case when you use blitz time control and monarch is clearly better than faile at that time control. Movei is losing against tscp if I tell it to play 5 times faster at 1 minute per game on p800. If I do the same with 10 minutes per game then movei is clearly winning. Imagine an engine that is 5 times slower than movei. It is going to lose against tscp at bullet and win against it at longer time control. Uri
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.