Author: Robert Hyatt
Date: 12:45:52 08/24/01
Go up one level in this thread
On August 24, 2001 at 09:48:39, José Carlos wrote: >On August 24, 2001 at 09:19:23, Robert Hyatt wrote: > >>On August 24, 2001 at 04:31:27, Tony Werten wrote: >> >>>Hi all. >>> >>>I like to know if there are some solutions for this problem. >>> >>>My program searches for 4 minutes, gets to 14 ply (or whatever), makes a move, >>>takes a pondermove and starts pondering. >>> >>>Now the first 12 ply are taken from the hashtable (<1 sec ) but it starts an >>>(almost) uninformed 13 ply search from which it is never going to return in the >>>allocated time, so it's just taking the best move it had from the previous >>>search, which is costing me plydepth. (and games ) >>> >>>Any ideas, sugestions ? >>> >>>cheers, >>> >>>Tony >> >> >> >>Internal iterative deepening can help here. The problem is that if your >>first move comes from the hash table with a good score (typical on pondering) >>then you reach depth N with a PV that has only that move in it. (there are >>things you can try to do, such as extending the PV in such cases by probing >>for the second position, etc when you back up the PV). IID works by doing >>shallow searches to find good moves at PV-type positions where you have no PV >>move. If you don't do something like this, you reach critical positions with >>no real hope of having the best move searched first, which makes the tree >>explode in a bad way... > > I'm sorry Bob, I don't follow. What's the difference between that second >search two plies shallower and adding a new ply to the main search and making >the first two moves of the PV? You should have the same information in the hash >table, assuming you didn't overwrite it in the ponder search, which is very rare >since you're hitting the hash table up to ply 12 (in his example) and not >searching anything down here. You have to stop and think about the math a second. You are about to do an N ply search for a position where you have no idea which move is best. If you do a N-2 ply search, that is _far_ less effort, and it gives you a move to try at N. A better move than just a wild guess. And move ordering is critical on alpha/beta, particularly along the PV. > So -sorry for my difficulty to understand-, why don't you have the PV in the >hash table in that ponder search? It can get overwritten. The first few moves ought to be there, but the last half can easily be overwritten after those nodes are searched... > > José C. > >>You can see how I implemented IID in Crafty. It is actually pretty simple, >>so long as you have an easy way to search the same position twice without >>causing your data structures any sort of problems..
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.