Author: Sune Fischer
Date: 02:51:54 03/18/04
Go up one level in this thread
On March 18, 2004 at 05:01:24, Mikael Bäckman wrote: >On March 17, 2004 at 18:12:45, Sune Fischer wrote: > >>On March 17, 2004 at 16:14:40, Mikael Bäckman wrote: >> >>>I used 90 seconds per position as I didn't know how deep I could search without >>>spending days on this... First I ran a test without historytables, to get a >>>depth to compare the other tests to. Most of the depths were completed in 20-60 >>>seconds. Perhaps a bit shallow, but it gives an idea of the performance. >> >>I'd prefer fewer positions and deeper searching. >>The global table only suffers a mild saturation in a shallow search, to really >>see the effect it must saturate badly and that takes a longer search ( > 100M >>nodes. >> >>>I use a side-piece-to historytable or history[side][piece][to] and I use at most >>>8 history moves at a node. After that I try the moves in the order they are >>>generated. >>> >>> >>>Test1 = No History >>>Test2 = History >>>Test3 = History - root aging >>>Test4 = History - age as soon as a history score gets larger than 10000. >> >>10000 didn't work for me. I think it is too aggressive, you 'age' the table 10 >>times a second at this rate. Try with a larger number like 65000, that's about >>once a second. >> >>You don't want to throw the baby out with the bathwater :) >> >>-S. > >Before the longer test, I ran a few positions for about 10 seconds and looked up >the largest value in the history tables. A value of 10000 seemed to give me >about one aging a second. Everything revolves around this parameter, if the parameter is too small it does not work, it will probably even be worse. If it's too big it is not optimal but still works, so approaching from the top is safer. >Maybe we update the history tables differently? I use history[] += d*d, where d >= depth/FULLPLY. FULLPLY is 8. A long time ago I tested this and I got slightly better results in testsuites with just += d. I don't really have a logical explanation for that, perhaps it is time to test it again. -S. >/Mikael
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.