Author: Ricardo Gibert
Date: 03:25:27 04/23/02
Go up one level in this thread
On April 22, 2002 at 09:01:30, Jouni Uski wrote: All 3 appear true to me. >1. program X plays relatively better against humans than against computers This can be illustrated by considering 2 possible programs: Program A tries to play open positions. Program B tries to play closed positions. When either program plays another program, they play on a more or less equal footing. For example, in closed positions, they are equally clueless. When Program A plays humans, it does reasonably well, since it tries to play positions programs are good at against humans. When Program B plays humans, it does poorly, since it tries to play positions programs are poor at against humans. This is, admittedly, an artificial example, but how it differs from actual programs is only a matter of degree to me. >2. program Y plays relatively better with longer time control This should tend to be true of programs with a lower EBF. Naturally, this presupposes the program has arrived at its lower EBF by some reasonable means e.g. a program uses R=2 nullmove in non-endgames, while an otherwise identical program does not. >3. there are diminishing returns from speed doubling, when search depth is > increased In theory, you can double the speed until a program plays perfectly. Further doubling will not improve the programs play. Twenty years ago, the rule of thumb was that programs increase in strength by 100 rating points with each doubling of speed. Today the rule of thumb tends to be more like only 40-50 rating points. It's gradually coming down. Not conclusive, but it would be pretty weird if "3." were not true. > >I think definitely 3. is true - only the size of diminishing is unclear. E.g. >from latest SSDF list (yes comp-comp play!) doubling gives only 40 points at >current top level against 75 previously. But 1 and 2 are still open cases. > >Jouni
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.