Author: Gabor Szots
Date: 04:09:38 03/31/00
Go up one level in this thread
On March 31, 2000 at 02:05:45, Dann Corbit wrote: >On March 31, 2000 at 01:34:22, Gabor Szots wrote: > >>Once I made experiments on how hash table size affects playing strength. I used >>BS-2830 test and LG2000 (I don't remember the version) for the purpose. On a >>Pentium 166 MMX I found that increasing hash table size from 4 MB to 8 and then >>to 16 increases Elo rating by 10 and 15, respectively. Increasing it to 32 >>probably would not have had any measurable effect at all. >>I know it also depends on how fast the machine is, what program is used, what >>time control is etc. but there must be some guidelines as to how much extent it >>is worth increasing hash at the cost of reducing memory size for other programs. >>By the way, can someone tell how playing strength is affected if I for example >>play flipper in the background (single processor machine)? > >Depends on your purpose. >The increase in playing strength looks like a logarithmic curve. As you >continue to add the same increment of hash memory, the added strength is not >nearly so much as at first. But suppose you are playing a contest. And >doubling hash from one hundred megabytes to two hundred megabytes adds 2% >playing strength. You will still do it because you want any edge you can get. >But if you want to be able to run your word processor and other stuff at the >same time and it is not for a contest then you won't care nearly as much. It looks I could not express clearly what I meant. I think there must be an upper limit for hash above which you can't increase playing strength. This is why: There is an upper limit of nodes a program can evaluate within a given time interval. Consequently the amount of data put into the hash during this interval is also limited. If you allocate too much memory for hash, some part of this memory remains empty, since the engine has no time to fill it. This memory is lost for you. Therefore, depending upon the speed of the machine, the time given for one move and the complexity of the evaluation function, it makes no sense to exceed a certain size. I am looking for an approximate formula with which I can determine this size. Yours, Gabor
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.