Author: Robert Hyatt
Date: 22:16:54 05/12/01
Go up one level in this thread
On May 12, 2001 at 01:20:59, Angrim wrote: >On May 11, 2001 at 17:16:26, Dann Corbit wrote: > >>On May 11, 2001 at 15:49:19, Angrim wrote: >> >>>On May 11, 2001 at 03:29:43, Dann Corbit wrote: >>>>Which brings up another thought. What percentage of moves are so horrible that >>>>they are not even worth considering. Is it 99.99999999999999999999999999%? >>> >>>not of moves, but of the set of all possible games, the percentage that contain >>>an error of that magnitude is roughly 99.999<insert 2 pages of 9s>99% >>>Even if you define such an error as "any move which can be shown to lose >>>with a 1 second search" rather than the possibly unsound "any move which >>>crafty would score as 10 points lower than the favorite after a 1 second >>>search". >> >>I don't think that such a percentage can really be quite so high. I notice in >>the middle game that there are usually about 35 choices that can be made. Super >>GM's will usually pick about 3 different options in such an example. So that's >>about ten percent from a given node. Now, it is interesting that it is a sort >>of exponential trimming, since (.10)^k grows towards zero quickly. But then, >>that's just another way of saying that the average branching factor is about 3, >>which means we are right back in exponential-ville. >> > >a 90% prune rate is much more optimistic than I was thinking. >Even allowing for the fact that "random" positions will generally have >much less king defenses than positions reached by solid play, I doubt that >more than 50% of the legal moves would lose trivially. >However, if you estimate that most of the legal games are at least 5000 >moves long, and 50% of the moves at each ply are errors, then the fraction >of those games which include an error would be 1-(0.5^5000) which has a >lot of 9s in it :) > ><big snip of factoring stuff> >wow, you really are bored. Whats the address for the >factoring message board? :) > >Angrim One simple way to test this hypothesis would be to take a real game or two, and feed them to crafty's annotate command, telling it to produce the score/PV for _all moves_ at each move in the game. See how many are going to change the outcome from win to draw or draw to loss. That would give some idea...
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.