Author: Robert Hyatt
Date: 10:03:35 09/05/01
Go up one level in this thread
On September 05, 2001 at 12:46:29, Adam Oellermann wrote: >On September 05, 2001 at 12:15:37, Uri Blass wrote: > >>On September 05, 2001 at 07:53:50, guy haworth wrote: >> >>> >>>In the Ken-Thompson-themed ICGA Journal (June, 2001), Ernst Heinz published his >>>latest self-play experiment results. >>> >>>Engines with different guaranteed-depth(?)-parameters were pitted against each >>>other. >>> >>>The matches of the experimetn (3,000 games each) suggest that: >>> >>> 12-ply was 84 ELO points better than 11 ply >>> 11-ply was 92 ELO points better than 10 ply >>> 10-ply was 115 ELO points better than 9 ply >>> >>>Fairly strong indications of decreasing returns from increasing search. No >>>doubt a proper statistical analysis will follow. >>> >>> >>>An extra ply seems to require 4-6 times the 'effective power', so a factor of 36 >>>- if realised across the system - is only 2 plies. >> >>I know that the top programs of today have usually branching factor that is >>close to 3 and not 4-6 so a factor of 36 is more than 3 plies. > >Ernst Heinz probably used Dark Thought for this research - while perhaps not a >top program these days, I reckon it's branching factor must be comparable to the >top programs. The article didn't say that the branching factor was 3, but that >the "effective power" (I'm guessing CPU time and memory utilisation are factors >in "effective power"), determined empirically, required for an extra ply is 4 - >6. This is wrong for today's programs. Don't know why his takes 4-6, but that is very _high_ today. 3 is a much closer number. >> >>I also know that based on the ssdf results 70 elo per doubling the speed makes >>more sense. >> >>70 elo per doubling means something like 110 elo per ply. >>84 elo per ply when going from 11 plies to 12 plies seems to be wrong because >>the programs in the ssdf games get deeper than 12 plies. > >He's not saying 84 elo per ply; he's saying that once you hit 11 ply, the next >ply will buy you around 84 elo. > >>Uri > >I think discounting the results published by Ernst Heinz because of the SSDF's >figures is dangerous. The SSDF, after all, is an experiment to test relative >performance of different chess engines. Ernst Heinz's experiment is clearly >designed to determine the performance improvements obtained by increased search >depth, and in this domain his data are probably more accurate than SSDF figures. > >- Adam
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.