Author: José Carlos
Date: 09:54:59 03/29/01
Go up one level in this thread
On March 29, 2001 at 11:22:49, Robert Hyatt wrote: >On March 28, 2001 at 15:31:57, Christophe Theron wrote: > >> >> >>Last time I checked it was 70 elo for doubling, and I have computed these >>figures from the SSDF list (not the latest however). >> >>Some people say 70 will not be sustained for higher and higher speeds (depths), >>and if it is the case then the gain for a dual is going to be less than what I >>have written. >> >> >> >> > > >I think that is an interesting thing to discuss. IE so far, I have not >seen _any_ evidence that says 2x faster today doesn't help as much as 2x faster >a year ago. Many "feel" that there must be some sort of asymptotic behavior >but so far there is no proof. In fact, there is ample evidence that for the >next few doublings, we will continue to get just as much improvement as we >did in the past. Newborn wrote a paper discussing this. Heinz followed up >with a verification experiment. Both show that deeper searches will _continue_ >to improve playing skill for a good while yet. I am one of those who "feel" that the behaviour must be asymptotic. Actually, I have absolutely no proof to support my feeling, but at least, some reasoning: - Doubling speed is, on average, less than an extra ply, given the actual branching factors. So, as the tree grows faster than we go deeper, we are deepening more slowly every time we double speed. - Even if we could achive a bf of 2, so that doubling speed would give a constant ply deepening, we could discuse if a ply add is worth the same at ply 7 or at ply 40. Again, no proof exist, but I tend to feel that something that is 40 deep in the tree has less influence than something that is very near, specially considering that, if I search 40 moves deep, probably I'm gonna win every game, so it doesn't matter much if I see 40 or 41; I'll win anyway. >yes, it is very probable that 20 years ago a doubling would be worth more >than 70 points. But it seems to have stabilized to around 70 for a long >while now. Yeah, it's a normal behaviour of an asymptotic curve. It's slowly flattening. >And no, I am _not_ talking about each doubling walking the computers up the >FIDE rating list by 70 points. I doubt doubling is worth much at all when >playing humans in the top 30 or so. But in pure comp vs comp games, give >me that factor of 2x every time. It will swing the balance in my favor for >a long long time I think. This is what I've said for some time. IMO, in comp-comp, small differences in playing strength make big difference in rating points. It seems that the scale is different. That's why it is impossible to compare ssdf and fide ratings, unless without doing some calculations. It's something like celsius and farenheit degrees: different scales, different starting points. If we could take a given program's ssdf and fide rating (in this second case, based upon a big number of games against humans) and the same with another program, we could extrapolate the fide rating for any program between them both in ssdf (I hope my bad english is not bad enought to not being undestood). Anyway, chess is much more complicated than degrees transformations, so we could only expect an approximation, and we'd need a big number of games. José C.
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.