Author: Mark Young
Date: 12:21:54 06/10/98
Go up one level in this thread
On June 10, 1998 at 11:59:37, Christophe Theron wrote: >On June 09, 1998 at 23:30:30, Mark Young wrote: > >>On June 09, 1998 at 22:52:35, Christophe Theron wrote: >> >>>On June 09, 1998 at 21:10:24, Mark Young wrote: >>> >>>>On June 09, 1998 at 17:02:51, Ed Schröder wrote: >>>> >>>>>>Posted by Christophe Theron on June 09, 1998 at 14:01:59: >>>>> >>>>>>>My view, in total 30 >>4:30 vs 45:00<< games will be played. Imagine >>>>>>>the following possible scores: >>>>> >>>>>>>a) 4:30 vs 45:00 5 - 25 --> Speed is decisive. >>>>>>>b) 4:30 vs 45:00 10 - 20 --> My expectation. >>>>>>>c) 4:30 vs 45:00 13 - 17 --> Speed is not decisive. >>>>> >>>>> >>>>>>I you use the top programs currently competing in CCL, my bet is that >>>>>>you'll get the 5 - 25 result. >>>>> >>>>>It's an old discussion. My view is that going from 5 to 6 ply is a lot >>>>>more worth than going from ply 12 to 13 and going from 12 to 13 ply will >>>>>gain more elo than going from ply 18 to 19. >>>>> >>>> >>>>From the testing I have done in the past. I think that this view is most >>>>probable correct. >>>> >>>>>IMO there comes a time (say about 10 years?) that a doubling of >>>>>processor speed will gain only 2-5 elo points. How much worth is >>>>>a doubling these days? A NPS tournament can reveal something about >>>>>this. >>>>> >>>> >>>>This might be true running todays programs, but you programmers are very >>>>good at finding new ways of using spare CPU horsepower. >>>> >>>> >>>> >>>>>My speculation is 10-20 as the current top programs at such fast >>>>>machines are already so good that they can't be slaughtered with >>>>>your suggested 5-25. >>>>> >>>>>- Ed - >>>>> >>>>> >>>>>>This just will tell us nothing. I suppose you think differently. So can >>>>>>you please explain what kind of lesson you are expecting from this? >>>>> >>>> >>>>It’s amazing just how little the creator truly understands the nature of >>>>the monster he makes. >>> >>>Maybe I don't know my creature very well, but you can be sure I know it >>>better than anyone else on earth. >>> >> >>I did not mean that comment to be a slam against you or anyone. > >No problem, Mark. > > >>I just >>think it's interesting how much we still don't know. If we all knew the >>out come of this testing, then it would be pointless. > >Of course we don't know the outcome of ANY test before doing it. > >It's just that I see lots of complains from people saying: "the SSDF >should always use equal platform for testing". And now somebody wants to >do a very unequal platform test and people say: "Oh yes, great, let's do >it, we have been waiting for that!". > >And I still don't understand what you are expecting from this test. > I’m not sure what to expect. I have worked on a projection of how I think the match may turn out. I took the data that I have on NPS matches that I have run, and tried to use this data to project the score of this match. At 10x this is about 3 and ½ doublings. I needed to know how much 1 doubling was worth on a P II 266, my data showed about 45 rating points. Then I calculated what the next 2-½ doublings should give me. So my data suggested 45 rating points for the first doubling, the second doubling would give me 33 rating points, and the third doubling would give me 24 rating points. The remaining part about 15 rating points. So my very limited data suggested a ratings gain of 117 points at 10x. So out of 30 games the 10x side should win 19.8 games. So the match should end 20-10. I think this is a very interesting test. That should be run at least a few times to confirm any data of previous results. I do agree that there is other kind of testing that needs to be done. And I’m sure it will be. >I think there are more interesting tests to do. For example, let's start >by making a few test matches with a 2x time handicap before trying 10x. > >Or even by giving a 20% time handicap. Sometimes I wonder: "should I try >to implement more knowledge, or should I spend the next month trying to >improve Tiger's speed by 20%?". > >Some months ago, Tiger was crushed by Genius in every blitz game. I was >really disappointed by the result. Then I tried handicap games, for >example by giving Tiger twice the time Genius had. Suddenly, Genius >looked rather weak. > >With this experiment I was able to know that it was worth keeping >working on my program. Because I knew the goal was near. > >I would have learned nothing with a 10x time handicap. When you give one >side too much handicap, this side just looks stupid. 2x time handicap is >already too much. 10x is not worth trying. > >Just my opinion. Let's see what happens, and then we talk about it >again. > > > >>>Gives you an idea of the kind of understanding most people (including on >>>CCC) have about chess programs. :) >>> >>> >>> >>>>If we knew what lessons we would learn for running this experiment there >>>>might not be a need to run it. >>> >>> >>>You are right, that's why I added the following comments (see quote >>>below): >>> >>> >>>>>>But OK, as I said, this has to be done at least one time. >>>>> >>>>>>And BTW I hope to be wrong. I like to be surprised! >>>>> >>>>>> Christophe
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.