Author: Robert Hyatt
Date: 08:48:22 09/06/02
Go up one level in this thread
On September 06, 2002 at 09:04:43, Vincent Diepeveen wrote: >On September 05, 2002 at 11:52:20, Robert Hyatt wrote: > >i don't care for icc. > >I hope you realize if tree shape of 2 processors is the same >and overlap is better of the tree, that speedup is better. That is, as we say, "intuitively obvious to the casual observer." > >That's very trivial. It is the reason why searching deeper >tends to give a better speedup for DTS. Because the >hashentries give you directly a similar shape then. > >Or do you deny this? Your reason is _wrong_. The reason deeper searches give better results with DTS is about _choice_. Not "better tree shape." All about "choice". If I do a 5 ply search as I did in my dissertation, when I do a split, I have at most 5 places where I can do a split with any one processor. If I do 12 ply searches as I am doing with Crafty, I would have 12 places to choose to do a split. That gives me more _choices_. And therefore a better chance to find a "good" choice. With five, I often have to take what I can get. with 12, if you read the DTS article, I could do some analysis at each of the 12 plies, and choose the one with the highest probability of being searched completely. _that_ is the benefit of more depth. If just going deeper would help, I ought to be able to search kopec2 to depth=10, and then to depth=12, and get a better speedup at 12. That doesn't happen. Because in Crafty I don't get to make a "choice" of where to split, I take what comes along by serendipity, because of the sorta-YBW algorithm I am using due to the recursive search. that's all there is too it... I think a deeper search might help in general, for positions where nothing unusual is going on. Because as you go deeper, move ordering gets better and better near the root, because of hashing and history counts and the like. But that is just one _class_ of positions. There is the other end of the spectrum where each new iteration reveals some new tactical issue that breaks all the old move ordering down, because each time you go a ply deeper, you discover another tactical issue that negates everything you have been computing so far, ordering info and all. Both are equally probable in a game. And it doesn't have to be a tactical position where you win or lose material. If you understand an outside passer but don't understand a distant majority, then if you search positions where there is a potential outside passer, your search can continually find a good pv at depth N, but at depth N+1 find that the opponent can produce that outside passer, but you find a move to prevent it. But the next iteration you discover that move now is not good enough, etc... that score change might only be .3 or .5 pawns, but it is more than enough to totally wreck the shape of the tree and even cause a serial search to have a gross branching factor.
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.