Author: Stuart Cracraft
Date: 11:34:48 08/09/04
Go up one level in this thread
On August 09, 2004 at 10:34:43, Jan K. wrote: >my opinion:see should probably help you even if you don't use it for qsearch >pruning, just for move ordering only. And even if you prune qsearch in the most >simple way (see(move)<0) it must prove to be a big win. Do I guess right that It should. But it didn't. >those 300 position are WAC test? The idea is that you can't assume anything >about improvement see gave you after only running wac or some else "5seconds to >go" test. Running the test simply doesn't give you any relevant information, you These are WAC. Yes -- I don't see why a test is necessarily bad. I'm just trying to keep bugs from getting in after changes. >may only find some bugs, if you get too many bad results. And the fact also is, >that with this short time test, how much nodes do you think you can prune? Even >with 50% smaller tree thanks to see, will it help you? No, you will search 1 ply >deeper and still have bad results. If you want to find the best way to prune >qsearch, you will have to try what works best for you, nobody can help you with >that. At the deeper searches the see()>=0 code gives 50% fewer nodes and 50% less time for the same deeper plies. I don't reach these in the 1 second tests (WAC) I do now on this 1Ghz P3, but in fairly short order, the same 1 second test will be searching deeper due to new hardware so the see() improvement will reflect even in short tests. These kinds of testsuites are invaluable since I can tell if the time, the nodes, the average depth, the total % correct, etc. are getting better, worse, or staying the same. I will not improve a program based on theory alone and games with me. I am too poor a player to go that route. Thanks, Stuart
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.