Author: Robert Hyatt
Date: 11:36:32 11/16/98
Go up one level in this thread
On November 16, 1998 at 12:09:33, Amir Ban wrote: >> > >I don't think so. This one is not a "nonsense" PV. It's genuine and needs an >explanation. > >Amir Here's a very precise question for you: Have you *ever* had Junior analyzing games where others could see it's output? An example: I often have Crafty give online analysis on ICC when GM's are playing in major tournaments. And in *every* game, someone will do this: Take a PV returned by a 13 ply search, play the moves on a "scratch" board, then say "hey all you guys that were amazed that Crafty had predicted all of the last 30 moves but one, look at this position after it's analysis." And invariable there will be some funny stuff at the end of the PV that produces a position that is quite questionable when looked at in that light. So, the point is, then, is it really important to look at moves further down the PV and try to justify *them* as well? Or is the move played at the root the place where we want the most accuracy? I'd *love* to do 13 ply searches and produce 16 move PV's where every move was of GM caliber. But I know that the root move will be *very* good. The second move a little worse because it is a 12 ply move... the third move a little worse still... etc... I often see oddball moves near the end of a PV, when the goal is to simply prevent some "threat" that I analyze from happening (IE Qa2 to prevent your playing Rb2, even though Rb2 is totally unimportant while Qa2 puts the queen in a horrible position.) I fix what I can, but I still see this. In fact, I see this in *all* programs that I watch where the operator will cut/paste analysis and send it to ICC... I'm much less worried about what's happening at the *end* of DB's PV's than I am interested in what is happening at the *front*. I'm sure you must have seen positions where there was a sequence in the PV that goes like this: 1. Nc3 Bf2+ Kh1 or 1. Kh1 Bf2 Nc3. Same position, right? Not quite. That + drives the PV one move deeper along that path, so if there is something I can do deep in the tree I don't move my king, knowing he won't check to extend my search so that I can see the result, or if it is something bad, I'll move my king so he can't check me and drive the search deep enough where I see the bad thing. Horizon gets us all from time to time, still. Lots of reasons to explain an unexpected queen move. From horizon, to eval quirk (defending against a useless threat) to a parallel search anomaly to a hashing anomaly to an outright bug. But I don't see why "it needs an explanation." Maybe we are curious and we have to figure out why... but I certainly don't intend to spend the rest of my days explaining every odd thing my program does... otherwise I won't ever get anything else done, because it does *lots* of odd things still.
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.