Author: Robert Hyatt
Date: 21:03:21 07/22/01
Go up one level in this thread
On July 21, 2001 at 18:49:28, Miguel A. Ballicora wrote: > Before you over-analyze this, let me explain what is going on... If you could somehow search a perfectly-ordered game tree for this position, it takes 26 plies to win a pawn. Nothing you can do to shorten that distance. If you search a less-than-perfectly-ordered tree, you can notice the following oddball behavior: Suppose you are going to search to a depth of 20 plies. And suppose for the first 10 plies, white plays perfectly, and black plays poorly, so that after those 10 plies, it only takes another 10 plies to force a win of material from this 10-ply position. Let's call this P. Now you continue searching and you now try better moves for black. Good enough moves that you now discover that you can't win a pawn. But with luck, after (say) 20 plies, you can reach position P. And with the hash entry for P showing you can win a pawn, you realize that _this_ path wins the pawn by force. But the search is really far deeper than 20 plies... it goes to depth=30 in fact... because entry P was a 10-ply search result, and you grafted that on to the end of a 20 ply search. If you searched everything in the best-possible order, this could not possibly happen and it would always take you 26 plies to find the pawn win. But with this poor move ordering, you can actually gain plies. But only if the key positions manage to stick in the hash table and get grafted to the right tree positions. In short, you can actually improve your program, and yet make it take longer to solve fine 70. And you can make your program actually worse, but see it solve fine 70 faster. In short, don't pay much attention to the depth at which you solve it, so long as you don't go beyond 26 plies to see it. The shorter your search to find Kb1, the worse your move ordering is in that position... Finding it at 26 plies would be the best possible result, IMHO... most of us find it at smaller depths, however, for obvious reasons...
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.