Author: Robert Hyatt
Date: 12:12:59 11/26/01
Go up one level in this thread
On November 26, 2001 at 01:23:02, Tom Kerrigan wrote: >On November 25, 2001 at 21:31:04, Robert Hyatt wrote: > >>No, I didn't say neither approach is capable of finding "mistakes". You are >>again simply missing the point. We are talking about mistakes that the program >>can _not_ find on its own, starting a search at the point where the original >>real mistake was made. As a result, the program will "report" the mistake > >Since when are we only talking about these mistakes? Which post am I supposed to >re-read? I've been reading this one: This is the _only_ kind of mistake that makes any sense, when you think about it. Otherwise _either_ approach works just as well. The original idea was that searching back-to-front lets the program pass information back up the tree so that "problems" are reported sooner in the game, rather than being reported when the search is actually able to find the problem with no help. > >http://www.icdchess.com/forums/1/message.shtml?198458 > >and it doesn't qualify the word mistake anywhere. Instead of calling my reading >skills into question, how about you be a little more careful when presenting >your points, eh? > I didn't start the thread. I simply responded that I had tried annotation _both_ ways. I originally did it front-to-back, then after someone suggested the reverse, I implemented it and left it that way for quite a while. And I kept getting questions about "why does changing the hash size" greatly change the output for the annotate command?" and so forth. And I decided that the inconsistency was not worth the effort... that's _all_ I have said. That back-to-front introduces random inconsistencies that are (to me) undesirable. >As for your whole consistency argument, you are working from the fundamentally >flawed premise that if a shallow search can't identify a mistake, then a >somewhat deeper search can't either. No idea where you get that idea. The topic being discussed is that a human makes a mistake at (say) move 25 in the game. He annotates the game using a computer and the computer reports that at move 35 he made a bad choice, and that by playing move X he produced a score of -3. Going front-to-back you won't see any negative scores until you get to move 35 (the move where the computer can see deeply enough to realize material is being lost). If you go back-to-front, that bad score will be propagated back up the tree to somewhere between move 25 and 35, where the program will report find a problem at a different point. And where this point is varies with the size of the hash (as the critical positions might not survive long enough to be used back up the tree) or the time of search (where critical positions might be overwritten due to larger search space). >Let's say you make a mistake at move 10 >that a shallow search can only identify at move 11. If you analyze >front-to-back, then yeah, you will consistently identify the wrong move as a >mistake. But if you analyze back-to-front, then there's a chance you will >identify the actual mistake as a mistake, right? So let's just get this >straight, you are saying that being consistent is worth more than increasing >accuracy? Some would beg to differ... > If it were one move, you would be right. But it typically isn't. It is _many_ moves. Say 10 moves or 20 plies. And the score doesn't get backed up the the _real_ problem move at all... >-Tom
This page took 0.01 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.