Author: Sune Fischer
Date: 03:08:02 03/17/04
Go up one level in this thread
On March 17, 2004 at 05:50:57, Mikael Bäckman wrote:
>On March 17, 2004 at 05:26:30, Sune Fischer wrote:
>
>>On March 17, 2004 at 05:17:43, Mikael Bäckman wrote:
>>
>>>>Just a single test position to depth 15:
>>>
>>>
>>>Testing on a single position is wrong.
>>>If you take 20 positions and repeat the test, are the results still the same?
>>>
>>>/Mikael
>>
>>True.
>>
>>I have tested on other positions back when I started using this, but as you can
>>see it takes a long time for each position so I just tried a single one this
>>time.
>>If the search is deep enough I always get the same pattern with the local
>>overtaking the global.
>>
>>It would be interesting to see if Crafty and other engines exibit the same
>>behavior, I would expect so but one never knows :)
>>
>>-S.
>
>I can run some tests when I get home, but I would need to know exactly how you
>'localize' the HH. Divide with X every N nodes?
Yes, pretty much.
I scale it down by a factor 8 when the entries hit a certain limit, ie.
if (hist[index]>X)
for (i...)
hist[i]=hist_seed[i]+(hist[i]>>3);
>My history tables in chepla look like history[stm][piece][to], it will be
>interesting to compare the numbers.
That should be possible, as I use the exact same type of table.
>One thing to check before running these tests.
>I had a bug a while ago where my History tables would overtake the good captures
>in movesorting: I add a base value of B for all good and equal captures. In some
>deep searches (mostly endgames searches) history values would become greater
>than B...
History scores can get quite big and tough to beat :)
One solution could be to search all the captures before you even assign history
scores to the non-captures.
-S.
>/Mikael
This page took 0.01 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.