Author: Robert Henry Durrett
Date: 16:46:54 05/26/02
Go up one level in this thread
On May 20, 2002 at 15:41:47, Eric Baum wrote: >On May 20, 2002 at 15:05:15, Russell Reagan wrote: > >>I would bet that a program with an evaluation function that primitive wouldn't >>break the 2000 elo mark, or maybe between 2000-2100 at best. Of course I could >>be wrong and someone like Bob or Dan Corbit could tell you better the >>differences between a program with a simple evaluation function vs. a program >>with a complex one, > >Hopefully one of the experts will respond :^) > >>There are programs like this, and they do learn new features. They generally use >>neural nets or something like that. They also generally stop improving at about >>a beginner level. > >I'm only interested in ones that actually do better... > >> >>>Also, for comparison, does anybody have a recent estimate of rating >>>point gain per additional ply of search? >> >>I don't, but someone does I'm sure. I would guess however that at some point you >>aren't going to get many more rating points, and then once you reach a really >>deep depth, you will start to see more jumping up of the rating, then another >>diminishing returns area, then another jump, and so on, until you reach a ply >>depth where you solve the game. > >Hsu used to give talks graphing rating increase vs ply and claimed something >like a 200 point increase per ply (I'm just going from memory, so number might >be in error). He used to claim, I think, that this increase would continue >to arbitrary depth. (Of course, he was trying to convince IBM management to >continue funding development of his parallel machine, but subsequent events >lent some credence to his previous prognostication.) > > >>And since I don't even know what "context dependent forward pruning" is, maybe >>you could explain that :) > >What I mean is some method of doing something like what humans do: >deciding to pursue some lines of search and abandon others, based on the >board position. When people talk nowadays about "forward pruning," are they doing that only at the start position prior to search, or do they also do this at later points in the search? Intuitively, it seems there would be a tradeoff between the time required to do additional position evaluations [later in the search] versus the improved quality of the rationale for additional pruning. In other words, a series of pruning activities might give the best evaluation tree following the original position to be evaluated. >Logistello, for example, does additional pruning beyond >alpha-beta in Othello, and the Deep Blue crowd tried singular extensions, >but again the singular extensions never added much to their ratings either. >These were forward pruning methods, but didn't really look at the board >position, really only looked at the evaluation of the board and evaluation of >other lines >to make a decision what to prune. So singular extensions prunes when one line >is much more highly evaluated than others, independent of what the actual >positions are, but my understanding is SI didn't lead to much gain. Is there any >serious forward pruning going >on, leading to real ratings improvements? And if so, is anybody gaining >from pruning based on position as opposed to merely numerical evaluation?
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.