Author: Robert Hyatt
Date: 20:09:41 11/07/97
Go up one level in this thread
On November 07, 1997 at 15:25:53, Willie Wood wrote: > >On November 07, 1997 at 14:36:07, Robert Hyatt wrote: > >>On November 07, 1997 at 13:22:10, Komputer Korner wrote: >> >>>I throw this out to all chess software programmers. Why can't the screen >>>show the alpha beta window at all times during the search as an option? >>>Surely this information doesn't change so much that it would be >>>unreadable. >> >>if you use PVS as I do, it will *always* simply be score, score+1, where >>score is the score you see for the best move so far. That wouldn't be >>very informative. if you don't use PVS, then yes it changes very >>rapidly >>as different parts of the tree are traversed and affect alpha/beta/both. >>perhaps thousands of times per second... > > >Bob, > >I could've sworn you said you didn't use a zero-width search, due to >inefficiencies caused by pv oscillation. Or does PVS (score, score+1) >not imply zero-width? Sorry... we get hung up in technical issues at times. I use PVS, which basically searches down the left-hand side of the tree with a normal alpha/ beta window of X+n and X-n, where X is the score from the last iteration and n is some aspiration window (I use about 1/4 pawn). The remainder of the nodes are searched with a null-window. At times, only a few nodes are searched with a non-null window, when ordering is very good. At other times, lots of nodes are searched with a non-null window due to poor move ordering. The post you are thinking of was relative to the mtd(f) algorithm, where *every* node is searched with a null-window. If you are off on your initial estimate, you have to research everything with a higher or lower null-window. In the case of Crafty, with fairly "healthy" positional scores, this caused a lot of grief, because the previous ply's search is not a very good estimate for the current iteration. For me, mtd(f) made the tree larger, while also making it harder to debug since you can't get a reliable PV in many cases. I've been using PVS since maybe 1980, with good results. First program I saw using it was a version of Blitz. Murray Campbell (as far as I know) developed the idea and mentioned it at an ACM event. I told him we could easily test it on our VAX. I modified the code in about an hour or so, and it looked pretty good, cutting the tree size way down. Later that night, our cray died, and we switched back to the vax as a backup, and I then discovered we were getting lots of fail highs. We got excited, but each time the score only went up by small increments, while in the normal program a fail high meant winning material. :) We got excited, but then realized what we were running. Murray had a big chuckle, and the algorithm worked perfectly. We used it from then on, as did Belle and many other programs of the era. It still works well for me... > >WW
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.