Author: Michael Henderson
Date: 10:12:54 09/17/04
Go up one level in this thread
On September 17, 2004 at 12:54:53, Scott Gasch wrote: >On September 16, 2004 at 22:40:35, Michael Henderson wrote: > >>On September 16, 2004 at 22:14:36, Stuart Cracraft wrote: >> >>>I have two variables "ttprobes" and "ttmatches". >>> >>>When I attempt a hash table probe, I always >>>ttprobes++. >>> >>>In the probe routine, if I get a match on the 64-bit >>>hash key, I do ttmatches++. >>> >>>My question is, for a tier-1 replace always table in >>>a standard null move R=2 searcher with material-only/pc-sq-only >>>evaluation, what kind of hit rate percentage should I expect with the >>>following calculation >>> >>> ttmatches*100/ttprobes >>> >>>Is there some other hash-table-related statistic calculation I >>>should be doing to see whether the hash table is contributing >>>its fair share to the search. >>> >>>Thanks, >>> >>>Stuart >> >>Usually it's around 20-30% in standard time control searches. 15-20% for hash >>cutoffs. I'm not sure about hash matching in 1 second searches but I would >>think it's relatively high because there is not much overwriting. >> >>Michael > >I'd say I usually see 30-40+% "matched the signature" and maybe 5%-10% "got a >cutoff" on my hash table. I use a 4 probe complicated replacing scheme for >storage. > >I just ran the initial position to depth 10 and got 32.4% matched the sig and >only 5.5% got a cutoff for the hash. > >In a hash-friendly position it can be higher... for example I just ran Fine #70 >to depth 30 and got 84.9% matched the sig and 48.1% cutoff rate. > >Scott It's kind of hard to answer Stuart's question without more info such as: 1. hash rate he is actually getting 2. hashing/probing qsearch 3. extensions/ no extensions 4. depth/ time of search/ position hash table works exceptionally well in opening and endgame, I'm surprised you're not getting more cutoffs from opening? If anyone has some stats on the hash cutoffs that would be great... Michael
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.