Author: Don Dailey
Date: 14:54:15 01/05/98
Go up one level in this thread
On January 05, 1998 at 02:58:14, Christophe Theron wrote: > >On January 04, 1998 at 18:37:30, Don Dailey wrote: > > >>Once Richard told me or Larry that he could not do a 2 ply search. I >>got the idea that his program always REQUIRES a response to the >>computers move. So a 1 ply search could never have a 1 ply PV if >>my understanding is right. And yet the search always seems odd except >>in the case where you are doing a 1 ply search with no selectivity. > >Don, I think you gave me a part of the secret last time. > >Remember that you said: > >>We could start with some simple experiments. Suppose we (as a test) >>threw out null move altogether and always took a beta cutoff at ood >>nodes, after ply 1, 3, 5, 7 etc. This of course would be when when >>the static eval >= beta. This first experiment would give us an uppper >>bound on how fast this could be. > >I tried this today. It doesn't work very well in the way you describe, >but I think you are on the right track. > >I changed your suggestion. I removed my quiescence search and turned it >to full width search BUT: >* if it is the opponent's move, I change alpha to max(alpha,eval(pos)) >before generating and trying the moves. This can produce a cutoff, as in >a standard QSearch, or give a better value for alpha. >* If it is the computer's move, I don't change anything >In both cases, I apply "futility pruning". That is: if >eval(pos)<alpha-margin, I don't generate every move, but only captures. >Let margin=0.90 >* I limit this Q-Selective-Search to 8 plies. > >To say it differently: I don't care if the opponent is threatened. And >if the score is very bad, I assume that only a capture can bring it back >into the window. > >This modified QSearch comes after a normal full width search. > >It doesn't work very well, because I only spent 1 hour or 2 on it. And >it is very slow too, almost unusable to play a real game, because it >develops large trees. > >But the interesting fact is that it gives Genius-like main lines. I mean >it solves some problems at ply 1, when a conventional program would >solve them at only deeper levels. In fact it solves only "traps", >positions where there is a danger for the side to move. And the main >line has very often an odd number of moves. Richards program acts like his full width depth with about 4 ply added, in others words if it says 4+12 he will see the same basic tactics most programs see in 8 ply. Your idea of "verifying stand pat" may be getting closer. I have the feeling you should be doing a 4 ply instead of 8 and then revert to ordinary quies. Just an idea! -- Don
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.