Author: Robert Hyatt
Date: 07:00:38 09/13/99
Go up one level in this thread
On September 13, 1999 at 04:37:58, Ralf Elvsén wrote: >On September 12, 1999 at 20:46:26, Robert Hyatt wrote: > >>On September 12, 1999 at 19:51:14, Ralf Elvsén wrote: >> >>>I have read about some experiments where you let a >>>program play against itself, but one version is searching >>>one or more ply deeper than the other. From this >>>you can estimate how much stronger a program gets >>>given more time to search. >>> >>>Is it possible to perform an experiment like this on >>>one computer according to the following? >>> >>>You take the principal variation given by the >>>computer which looks like >>> >>>move1 move2 move3 ... >>> >>>i.e. the computer plays move1 and assumes the reply will be move2. >>> >>>Then you consistently make move2 (just like when you're cheating >>>and look at the info from the computer) >>>in reply to move1. Move2 is based on a search one ply >>>more shallow than move1, so I figure this would work OK >>>to simulate the abovementioned experiments. >>> >>>On the other hand, both moves "aim" for the same position >>>(the one after the last move in the line). >>> >>>I haven't really been able to decide whether this would work or not. >>>Presently I take the position that this would indeed give >>>the same results as an ordinary experiment, but I sway back >>>and forth, so if you think otherwise, >>>you will not have a hard time to convince me :) >>> >>>Ralf >> >> >>It won't quite work. Because (at least for me) I treat things at the root >>differently than at ply > 1. For example, if I am in check at ply=1 I don't >>extend because that would extend _every_ move and it makes no sense. But at >>ply=2 it makes perfect sense to extend. This means that taking the first move >>off the PV and assuming that the rest would be seen by a 1-ply shallower search >>is not a good assumption... > >Ok, it won't work if you want to do a high quality experiment. >But I guess that in general move2 is one ply "weaker" than move1? >How much is a ply, a factor 3-4? That is about 100 rating points. >The effect should be quite easy to see. > >Ralf It won't quite be a ply... because, as I said, the first ply doesn't get extended the same way other plies do. Using the N-1 ply PV actually gets a better PV than if you do a real search to depth N-1 after making the first move...
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.