Author: Volker Böhm
Date: 10:55:29 12/07/04
Go up one level in this thread
Hi, I am currently testing hash for qsearch in spike. I had the following test results: 1. Reading from hash without storing to hash does not gain anything. The minor reduction in nodes searched is more than eaten away by the amount of time needed for the lookup 2. Reading from hash and storing to hash without using the hashmove leaded to about 10% node reduction, but took about 7% more time per node. I use a "two phase hash" with an replace if deeper and an replace allways part. For q-search I store only to the "replace allways part" to save time. Copying the implementation of hash probe/write from normal search leaded to 10% more time per node thus does not gain anything. 3. Reading from hash if eval does not fail high (before capture move generation) gained about 9% node reduction and took only 3% more time per node. Thus this is currently my favorite. Maybe this could be improved by checking if a "lazy-eval" could lead to a very fast calculated beta cutoff before checking hash. 4. Using the hashmove as best move (captures only) does not gain anything. The minor reduction in nodes searched ... (see 1.) I haven´t testet the result of qsearch hash in "real" games. I doubt that the gain in playing strength coud be measured in a usual limited amount of time. Greetings Volker
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.