Author: Stuart Cracraft
Date: 09:42:25 09/23/04
Go up one level in this thread
On September 22, 2004 at 02:51:55, Richard Pijl wrote: >On September 21, 2004 at 19:44:00, Stuart Cracraft wrote: > >>On September 21, 2004 at 18:01:26, Richard Pijl wrote: >> >>>On September 21, 2004 at 17:36:47, Stuart Cracraft wrote: >>> >>>>Hi -- this past weekend I switched from single-tier replace >>>>always to two-tier place 1st tier in 2nd if incoming position >>>>is searched to a >= depth than currently stored at hash entry >>>>and store incoming position in 1st tier, otherwise always replace >>>>2nd tier if depth is. >>>> >>>>This is represented by the actual code below. >>>> >>>>After doing this, I expected least the same result or slightly >>>>better (than 250/300 on Win-at-Chess). Instead I scored 248/300 >>>>(consistently) with Two-Tier and 250/300 consistently with One-Tier. >>>> >>>>I am asking that some of the talented folks look this over >>>>and tell me if this is grossly wrong (expectation -or- code.) >>>> >>>>Or what might be the best methods for evaluating the function of >>>>the two methods... >>>> >>> >>>Replacement schemes start making a difference when entries start to get >>>overwritten on a regular basis. So, don't expect a big difference on 1 second >>>searches. The overhead of a dual probe (without looking at your code) may hurt >>>at such short searches. >>>What you can do is reduce the hash size to, lets say 10kb and then compare >>>scores again. >>>Richard. >> >>I always wondered why aging of entries in the hash table using a timestamp >>wouldn't be better. >> >>Stuart > >I'm not sure what you mean here. Do you mean using aging of entries vs clearing >the hashtable between moves? >In the Baron I'm doing aging. The only reason I can find not to do this is >because the differences in distance to root of the search positions may cause >instability (e.g. because of preprocessing, distance to root dependent extension >limitations and pruning, etc.). >I see no reason not to use old entries for e.g. moveordering though. >Richard. Exactly.
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.