Author: Richard Pijl
Date: 07:33:21 09/06/04
Go up one level in this thread
On September 05, 2004 at 20:31:40, Michael Henderson wrote: >I take care of the PV during the search, but I have been wondering about the >other method of getting the PV from hash. Some of these are obvious >questions/answers but I would like to discuss them. IMHO There is no real reason to get the PV from hash. It may seem a bit more efficient, but the loss of information is not worth it. I changed to PV from hash some time ago. The Baron was a small engine with little knowledge. I did notice a few percent increase in nodespeed so I was happy. Later I had problems with getting short PV's. Changing back to remembering the PV's during search did cost me a little in nps, but I also noticed that the time to ply was less (in longer testruns) due to better moveordering. Currently I display the PV as remembered during search, and I add moves found in the hashtable (from previous search runs). To show the distinction, the moves from the hashtable are shown between brackets. Now to your questions: >How do you know the real length of the PV? You don't, but you can guess the minimal number of moves left from the draft >Do you not usually get short PV's due to a hash miss? It happens especially with longer searches, or a small hashtable. When I changed back to storing the PV in a separate array, I noticed a significant increase of PV length, and slightly better moveordering because of it. >If you don't know the "draft" (due to extensions) at any point in the PV how do >you know a hash hit with depth x is good? You don't. > >How does the path info problem with hash interact with the pv hash extraction >method? Take care you break off producing the PV when a repetition in the PV occurs >Does score backed up ever really go with the leaf position of the PV? Not often >I've heard this method can sometimes lead to blunders in the PV? I guess that when you break off the search before it completed its iteration, that can happen. It think it should not happen at the end of a completed iteration. >Shredder uses this method?
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.