Author: David Blackman
Date: 23:46:27 05/19/98
Go up one level in this thread
On May 18, 1998 at 11:28:14, Robert Hyatt wrote: >I tried it about 2 years ago... It does two things: (1) it will reduce >the size of the tree searched by about 10% according to the tests I ran >way back then. (2) it breaks programs that generate moves in batches, >because you can't... you have to generate *all* the moves, and use each >move to produce a new hash signature that you use to probe the hash >table >to decide which move to try. In Crafty, this cost me more than that 10% >saving in nodes searched. Suggestion: Divide the search into two (or more) routines. One for the quiescence search and the last 2 to 4 plies of the normal search, and one for the rest. The routine close to the leaves is optimised for speed. So it generates moves in small batches and it doesn't do anything that looks too expensive. This routine will account for maybe 99% of all nodes searched, but if it is a bit sub-optimal on things like move ordering, the effect on overall node-count is small. The other routine, further from the leaves, generates all moves before searching any. It does Enhanced Transposition Cutoff, and any other relatively expensive tricks you can think of, provided it is likely to reduce the node count. Deperado sort of works this way, and i will soon add ETC to the routine further from the leaves to see if it helps there. Certainly it won't cost too much.
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.