Author: Ernst A. Heinz
Date: 21:39:08 08/02/98
Go up one level in this thread
On August 02, 1998 at 12:54:49, Robert Hyatt wrote: > >one other note. If you are going to behave like this, you might consider >the "full" PVS implementation that at least two programs used in the late >70's/early 80's.. belle and cray blitz. > >when you fail high, mark that move as best but *don't* research. If a second >move fails high, you first research the first move with a relaxed beta value, >then you search the second to see if it fails high. If not, keep going and >stick with the first fail-hi move. If the second fails high, don't research, >just mark as best and see if another will fail high. > >What this really does is help those positions where A is best at odd plies and >B is best at even plies, because you never have to do a research. I didn't >like >it due to lack of something to ponder, and you won't get a best ply=2 move from >the hash table either. But I have alternative ways to find something to >ponder, >and with "internal iterative deepening" not having a PV at the start of an >iteration is not too harmful either... As a follow-up to the above, I like to cite a passage from the section "Search engine" of my article "How DarkThought plays chess" (ICCA Journal, Vol. 20, No. 3, Sept. 1997, preprint available from our WWW page at URL <http://wwwipd.ira.uka.de/Tichy/DarkThought>): "At the top level DarkThought employs lazy alpha bounding and iterative deepening with an aspiration window of half a pawn. In contrast to plain alpha bounding, the lazy scheme delays the complete resolution of both new best moves and fail highs up to the next iteration. Top-level alpha bounding often saves some effort while at the same time searching new best moves one ply deeper than usual. If the top-level search is unstable or the final score lies far below the score of the previous move, DarkThought extends the duration of the current search up to threefold when playing tournament mode." According to my own experience, the greatest disadvantage of lazy alpha bounding is the substantial increase in complexity of the top-level search. =Ernst=
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.