Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Evaluation Jumps in Gambit Tiger II and other misteries...

Author: Robert Hyatt

Date: 06:33:21 04/05/01

Go up one level in this thread


On April 05, 2001 at 01:28:10, Christophe Theron wrote:

>On April 04, 2001 at 23:23:02, Robert Hyatt wrote:
>
>>On April 04, 2001 at 19:09:19, Christophe Theron wrote:
>>
>>>On April 04, 2001 at 18:24:41, Robert Hyatt wrote:
>>>
>>>>On April 04, 2001 at 17:44:25, Christophe Theron wrote:
>>>>
>>>>>On April 04, 2001 at 15:20:08, Dan Ellwein wrote:
>>>>>
>>>>>>Christophe
>>>>>>
>>>>>>in regards to the following quote:
>>>>>>
>>>>>>"These changes are also the reason why I believe that Gambit Tiger needs a
>>>>>>little bit more depth than Chess Tiger to achieve its full strength. At very
>>>>>>shallow ply depths, there is too much uncertainty for Gambit. It needs more
>>>>>>depth to find stable king attack plans."
>>>>>>
>>>>>>Is it accurate to say that Gambit Tiger plays better at longer time controls...
>>>>>
>>>>>
>>>>>No, it is not what I'm claiming here.
>>>>>
>>>>>What I am saying is that Gambit Tiger is not suited for very slow computers.
>>>>
>>>>Sorry, but those two statements are the _same_ thing:
>>>>
>>>>1.  GT needs a faster processor to do ok;
>>>>
>>>>2.  GT needs more time to do ok;
>>>
>>>
>>>
>>>It's a matter of scale.
>>>
>>>I have made a few test games with Gambit Tiger 1.0 on a 20MHz computer, and it
>>>performed poorly because it was playing too obvious attacking moves that could
>>>be refuted just by looking a few plies deep.
>>>
>>>These were games in 10 minutes, and the program routinely reached 4 to 5 plies.
>>>
>>>On the other hand, look at the results of Gambit Tiger on current hardware at
>>>game in one minute (some results have been posted today) and you will see that
>>>this problem completely disappears very quickly.
>>>
>>>If you want to deduce from this that "Gambit plays better at longer time
>>>controls", very well. But it's of course not the case.
>>>
>>>And Gambit Tiger does NOT need a faster processor to do OK, unless you are
>>>talking about 5MHz computers.
>>>
>>>And Gambit Tiger does NOT need more time to do OK, unless you are speaking about
>>>0.05s per move.
>>>
>>>
>>>Just check the posted results if you want to be sure.
>>>
>>>
>>>
>>>    Christophe
>>
>>
>>OK... I'll bite on this discussion.
>>
>>1.  I do _not_ believe that you have a "magic depth" that once you reach that,
>>you don't need more.  If a slow machine doesn't search deep enough, while a
>>faster machine does, I don't believe that is a "unit step function".  I believe
>>it is a "continuous function" so that as the hardware gets faster, the
>>advantage continues to accrue.  Otherwise you could find some magic depth and
>>say "if I  search below this I get killed, if I search above this, I never
>>have a problem."
>>
>>2.  I have seen the _same_ problem in my program.  My aggression is not so much
>>directed at king safety, as it is directed at avoiding blocked positions that
>>human IM/GM players strive for.  But at very shallow search depths, it will
>>make mistakes that it can't defend later, while at reasonable time limits, it
>>will not make aggressive counter-moves that ultimately lead to a quick loss
>>of material somewhere.  There is a steady improvement as depth increases...
>>
>>I don't see why such an idea is bad or wrong either.  It seems intuitive to
>>me...
>>
>>And don't forget, you can double the time per move, or double the clock speed
>>of the processor.  The effect is _identical_.
>
>
>
>Maybe you are right, Bob.
>
>What I said is based on a few games I have played with Gambit Tiger 1.0 on a
>very slow computer, at a time control of game in 10 minutes. What I have seen
>convinced me to stop using Gambit Tiger on this kind of computer, at that time
>control.
>
>But I should maybe try again and maybe with enough games Gambit would eventually
>turn out to be as useable as Chess Tiger on slow computers.
>
>
>
>    Christophe


I run into real problems with null move and shallow depths, as I have reported
before.  This is one reason crafty is not a great bullet chess player.  Yes,
against humans it is usually murder.  But against a "safer" type of pruning, it
does very badly at 1 0 games.  It does better at 5 3 games.  Even better at
40/2hr games, and at 2 hours+ per move (ie Ed's 2010 experiment) it has no
problems at all that I could see.

I think the more aggressively you play, the more this becomes evident, too.  As
a "safe" program doesn't hang things out in the wind.  But an aggressive program
can be led to do so by the evaluation, unless the search refutes the idea.



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.