Author: J. Wesley Cleveland
Date: 11:57:49 10/20/99
Go up one level in this thread
On October 20, 1999 at 11:17:08, Robert Hyatt wrote: >On October 20, 1999 at 02:26:55, blass uri wrote: > >>On October 19, 1999 at 23:59:20, Jeremiah Penery wrote: >> >>>On October 19, 1999 at 17:15:30, blass uri wrote: >>> >>>>On October 19, 1999 at 16:57:41, Jeremiah Penery wrote: >>>> >>>>>On October 19, 1999 at 16:48:36, walter irvin wrote: >>>>> >>>>>>On October 19, 1999 at 10:30:18, rich buska wrote: >>>>>> >>>>>>>i understand that is a 1/2 move but what does equal in elo >>>>>>> >>>>>>>2 ply 1100 ? >>>>>>>4 ply 1400 ? >>>>>>>6 ply 1600? >>>>>>>8 ply 1800? etc? >>>>>> >>>>>>depends on program , but i noticed that you raise each ply by 200 .during the >>>>>>first few ply may be worth more than 200 . plys deeper than 7 are worth less >>>>>>usually .plys greater than 12 are worth even less vs humans .against computer >>>>>>200 might work out , but not vs humans. >>>>> >>>>>The available evidence disagrees. In "Crafty goes deep" and "DarkThought goes >>>>>deep" (JICCA, sometime), it is shown that every increase in depth is relatively >>>>>equal with regard to picking a new best move. Therefore, the ratings difference >>>>>should be roughly equal for each ply as well. >>>> >>>>It is not clear if the last consequence is correct. >>>>and there is no evidence that it is correct against humans because humans can >>>>go to positions when another ply is less important. >>> >>>But if the program picks a new best move when going to the next ply, it is just >>>as important. >> >>This is not clear. >>The question is if the new best move is leading to the same result of the game >>or to different result of the game. >> >>It may lead to different result more times when you go from x plies to x+1 plies >>relative to when you go from y plies to y+1 plies. >> >>Uri > > >I personally believe that a depth N+1 search _always_ produces a better result >(more accurate) than a depth N search, no matter what N is. Because even if the >two searches produce the _same_ best move, I know that it is more likely to be >the right move because one ply deeper still liked it as best. I _really_ worry >about those positions where one ply deeper shows that the current best move is >a loser. And if you look at a chess program's output for a game, how many times >does it fail low on the last iteration and have to change its mind? In each of >those cases, the extra ply saved the game, rather than allowing it to be lost >instantly... This relates to an idea I had for a new extension. Every time the PV changes, extend one ply (or possibly more depending on the search depth). I tried some experiments with Crafty and it appears that the PV does not change that often. Since the value the PV search gives to a position is the static evaluation of the one leaf node on the PV, I think that anything that would improve that would be useful.
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.