Author: Christophe Theron
Date: 09:44:40 04/15/00
Go up one level in this thread
On April 15, 2000 at 08:39:56, Laurence Chen wrote:
>Hi all,
> I've ran some blitz games between Junior6a and Rebel Tiger 12e, and Hiarcs
>7.32 and Rebel Tiger 12e, and I was surprised the results. I'm using two
>identical computers, Pentium III 600E, autoplayer 232, blitz games at 5 min. + 3
>secs. Junior and Hiarcs were using Fritz 5.32 opening book, and Rebel Tiger its
>default book with NSEW=1, MoreSel=1. In the match where all engines were using
>16 MB of Ram for hashtables, Rebel Tiger got trounced very badly, against Junior
>6a, it lost to a score of 15.5-34.5, against Hiarcs 7.32, it lost to a score of
>13-37. Jim Walker suggested that I use 8 MB of RAM instead, so I ran another
>match of Rebel Tiger against Hiarcs 7.32, both engines using 8 MB of RAM, and
>now Rebel Tiger lost to a score of 21-29. I'm sure there are dupes, but that's
>Rebel Tiger fault for having a bug in the learning book. Yes, I downloaded the
>patch for learning from Rebel's website, and it did not fix the problem. This
>clearly shows that giving more hashtable for Rebel Tiger in this time control is
>very bad. The engine performance goes down a lot. So this raises the question,
>what is the optimal hashtable for different engines at different time controls?
>One cannot assume that 8 Mb of RAM for Hiarcs is better, I will later on run
>another 50 games match with Hiarcs having 16 MB of RAM, and Rebel Tiger 8 MB of
>RAM. A lot of us think that more RAM increases the strength of a chess engine.
>From this little experiment, it clearly shows that the opposite is true. I
>believe that there must be an optimal hashtable size where the engine will
>perform its best when given the proper hashtable size for a specific time
>control. How would one figure this out? I think that test suites cannot
>provide a proper answer to this question, any ideas.
>Laurence
I have already answered this question several times, at least regarding
Rebel-Tiger.
At fast time controls, giving too much hash tables kills the program, because of
the time needed to update them at the beginning of each search.
If you give too much hash tables, Tiger will spend a significant amount of time
before each move in order to update the informations stored in the hash table.
Depending on the computer's speed, it can be 20% of the thinking time in the
opening phase, to 80% of the thinking time in the endgame.
In this case, the additional amount of hash table does not speed up the engine.
It actually slows it down badly.
In the next release of Rebel-Tiger, the engine will be equipped with an
automatic hash table size setting algorithm, which will choose the best amount
of hash table depending on the computer you are using and the time controls of
the game.
Until that, you might consider giving the engine less hash tables if you play at
fast time controls in order to optimize its performances.
Christophe
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.