Author: Robert Hyatt
Date: 06:18:05 01/15/98
Go up one level in this thread
On January 15, 1998 at 07:45:40, Robert Hyatt wrote: >On January 15, 1998 at 07:19:48, Amir Ban wrote: > >> >>Often when you evaluate a node, you have a situation where you can >>record a bigger depth in the hash entry, but the eval info is less >>exact. What do you do in this case ? >> >>As an example, suppose your best move was e4 up to ply 6 with value >>+0.10, and then at ply 6 you find d4 is better. At ply 7 d4 is still >>best with value +0.20. When you evaluate e4 at ply 7, you have the >>choice of updating e4 to 7/LE 0.20 or leave it as 6/EQ 0.10. There are >>advantages and disadvantages either way. >> >>Amir > >If you use the two-level approach, you can do both. IE, the approach >used by Ken Thompson in Belle had two tables, one "always" replace, one >depth-preferred. I have modified this slightly so that the >depth-preferred >table is the same, but the other does not get overwritten always. >Rather, >if the depth-preferred table gets overwritten, the entry getting >overwritten >is moved to the always-store table. Otherwise the new entry is written >to >the always store table, and the old sticks in the depth-preferred table. > >This means you *can* get two hits. I've never measured this in Crafty >to >see if the second hit provides anything useful that the first didn't. I >didn't implement this change for that reason, in fact. It sort of acts >(now) like a two-level cache. Depth-preferred is usually the best place >to get a hit, but the other is good also... One thing I forgot, that Ken's algorithm handles nicely, is that if you have a hashing algorithm that probes to one address, and you either replace or give up, you can choke and die on deep searches. The table can fill up with entries with deep drafts, and the shallow draft positions can't "get in". But they are critical when you are way out in the tree. And if you can't store them, the tree size blows up. Ken's algorithm always has a place to store shallow entries, because of the "always store" part of the table...
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.