Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Chess Tiger -- Hash dilemma

Author: Christophe Theron

Date: 15:59:24 08/23/00

Go up one level in this thread


On August 23, 2000 at 14:30:20, Keith Conary wrote:

>I currently have an active account which runs Chess Tiger automatically on ICC
>and I've found that running it with the suggested hash table increment of 4-8M
>works well for fast times against computers only.  But I get several draws and
>losses to IMs and GMs with the small hash table settings.  After increasing the
>hash table size to 64M, I get the opposite effect, meaning I crush the humans
>and lose to the computer at 5 3.  Is there a medium that can be reached in
>tigers' settings to balance the performance between humans and computers?
>
>Thanks
>
>Keith


Hi Keith!


As the author of the engine, I confirm that what you have noticed can only be
explained by "luck" (noise).

When you play fast time control games with Tiger, you must set the hash table
size to a low value (between 2Mb and 8Mb). This is because Chess Tiger 12.0
needs to clear the hash tables before every thinking process, and this takes too
much time if your hash tables are big.

With really big hash tables, you simply waste 0.2 to 0.5 second per move
clearing the tables, and after 100 moves you have simply wasted 50 seconds
clearing them, and will lose on time.

This problem will totally disappear in the next version of the engine (Chess
Tiger 13), and with the new engine the rule is simple: you give as much RAM for
hash tables as you can, even if you play game in 10 seconds. Because the new
engine does not need to clear the hash tables anymore.

But until you get the new engine, you need to set a low hash table size when you
play fast time control games.

And the hash table size has no effect on the playing strength against humans. As
far as I know.



    Christophe



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.