Author: Robert Hyatt
Date: 17:35:31 07/30/98
Go up one level in this thread
On July 29, 1998 at 23:45:09, Ernst A. Heinz wrote: >On July 29, 1998 at 18:54:35, Robert Hyatt wrote: > >>On July 29, 1998 at 13:21:05, Danniel Corbit wrote: >> >>>On July 29, 1998 at 10:48:44, Tom Kerrigan wrote: >>> >>>>On July 29, 1998 at 10:06:43, Danniel Corbit wrote: >>>> >>>>>Let's say, for the sake of jihad, that we define sensible moves in the following >>>>>manner: >>>>>From a given FEN position, look at what every GM who has ever visited that >>>>>position has done. >>>>>From the same FEN position, let every top commercial and amateur program examine >>>>>that position for 24 hours. >>>>>At this point, we would have, I suspect, a small list of possible very good >>>>>moves, a mid sized list of so-so moves, and a large list of bad moves. >>>> >>>>Perhaps. >>>>I don't know what this would accomplish, and it seems like similar things have >>>>been done by people who do opening book stuff. >>>>Basically, what you're writing isn't really related to the topic at hand. >>>I disagree. It is related, because I believe that an exhaustive search has very >>>little value, and a careful analysis of the "ahem" good moves has a very high >>>value. >>> >>>Here is an experiment that will demonstrate my point [I think]. Run any chess >>>engine you like against a set of 100 arbitrarily chosen FEN positions achieved >>>by at least 3 different GM's at least 6 moves into the game. Analyze the >>>position for 7 plys, exhaustive. Analyze the position for 8 ply's exhaustive. >>>See how often a move that was not considered one of the top three from previous >>>exhaustive searches gets introduced as the new choice. I suspect it will be >>>less than one in a thousand. >>> >>>If that is the case, then exhaustive searching has very, very little value. >>> >>>We will certainly not be able to search 12 plys exhaustive for a long time, >>>anyway. Exhaustive search will therefore never be competitive with Alpha-Beta >>>or any other real searching technique. So what is it's value? Only to find the >>>rare gem that conventional searching techniques might miss. 99.99999% of the >>>bogus games generated by the exhaustive search will be utter crap that even a 9 >>>year old novice would not play. >> >> >>Here I have some data. In the "Crafty goes Deep" experiment last year, we ran >>such a test with Crafty, searching 347 different positions to depth=15. We >>were interested in how often does one more ply produce a different move. The >>result? Somewhere in the 20% range which was surprising. In fact, within close >>limits, adding one ply changed the PV about 20% of the time no matter whether it >>was going from ply 6 to ply 7, or ply 14 to 15. The JICCA has the exact >>results, while the raw search output is on my ftp machine... >> >>Bob > >Bob, > >Your paper only mentions searches to fixed depths of 14 plies. > >=Ernst= actually wasn't my paper. :) Monty wrote the thing... I simply put the data together for him... However, looking at the raw data, almost all positions were searched to depth=15, but a couple were so complicated that after 24 hours they stopped at 14. That might be why the tables didn't go any further. I do remember that Monty told me via phone that Ken Thompson was most surprised that it was still changing its mind, from 13-14, about as often as it changed its mind from 10-11 or whatever... Which is one reason I will *always* take another ply when given the chance, right? :) Of course, some say "nps doesn't matter" but I notice that *nobody* ever plays an important game without trying to obtain the fastest machine that is available. :)
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.