Laskos wrote:Eelco de Groot wrote:
It is not a linear scale, the x-axis is more a logarithmic scale. What kind of statistics do you need for that? I do not know.
Eelco
A small correction, plies are logarithmic with time, Elo is also logarithmic, so Elo-plies is linear (assumed constant nearly 70 for two-fold increase in time). It is not actually constant, and Pade Approximants do a very good job when nearly asymptotic linear approximation is needed. Jarkko is doing a very good test actually, I hope he will succeed at 12 ply (at least) with 1000 games, if Bob doesn't come to give us 14-ply results. Then I could extrapolate finely. Trust me up to the ply 25 or even further. And we will see in several years if we were wrong or right. The first result is as follows: 70 Elo points for doubling the time as of now (fast game, 15 plies), 45 Elo points for doubling the time in 6-8 years (long game, 25 plies, with very much improved hardware). Would you bet on a bottle of Champagne Louis Roederer's Cristal on the first of August 2017?
Kai
If the bet is about a Toga that I can build myself in 2017, I would be willing to engage in
any kind of bet about ply levels correlated to my own Fruit derivative playing strength
The program would probably be called, still in 2017,
Blueberry, when I would still be capable of making chessprograms then, maybe some kind of fine winegrape from France would be appropiate especially for the occasion but not Toga I'm afraid, it certainly would have to be some kind of fruit!
I am not much of a champagne connaisseur though. A friend of mine treated me to a good glass of white, from a very small house, appelation controllee 2007 fine wine yesterday, unfortunately I can't quite remember the name of the house now. It was pretty strong, I had just one glass of it and it was affecting my coordinaton when talking a walk outside later. I guess I'm not used to it and it
was 13.5% alcohol. Mais la belle France! Go Lance!
I take your point about the double logarithmic axes Kai. But I still think there are factors that probably do not scale up logarithmically, at least not yet. If the hash table size as an example does not scale logarithmically with the size of the searchtree, for pure practical reasons, what effects does it have the quality of the analysis? It may be slightly
positive even to have smaller tables, if you have less errors from hash collisions etc, although it does not help in keeping the search times acceptable.
I am not sure also what positional errors in the searchtree do with greater depth, you assume downward progation is limited, so searching deeper is always better, but is it true when there are diminshing returns in general also? Is it possible to have
upward propagation effects of search errors, evaluation errors, bugs etc. that are
increasing somehow with search depth when the tree grows larger. I am totally speculating about this but I think you can't exclude the possibility
Just advocating any "other side" of a possible debate here although I am not sure what the "other side" is exactly taking position at yet
I would say though that, in my opinion, it is likely possible though that any extrapolation will asymptotically approach to zero, so zero extra measurable elo returns at great plydepth, when given the Toga 1.4 beta 5c used for Jarkko's experiment we did today, or same experiments on 2017 hardware, but the same program version, so that determines for a big part the general shape of your curve towards the tail I think.
I do not know if the asymtotical approach towards zero is an underlying assumption of the extrapolation method that you both chose? I have not looked up the extrapolation method that you used Kai, but it was mainly any presumed 10^-4 elo-accuracy of four places behind the comma of the extrapolation that I was questioning, firstly...
Regards, Eelco