Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: The Brick Wall

Author: Uri Blass

Date: 12:22:22 09/20/04

Go up one level in this thread


On September 20, 2004 at 13:17:50, Robert Hyatt wrote:

>On September 20, 2004 at 12:56:27, Uri Blass wrote:
>
>>On September 20, 2004 at 10:48:57, Robert Hyatt wrote:
>>
>>>On September 19, 2004 at 21:28:16, Uri Blass wrote:
>>>
>>>>On September 19, 2004 at 20:37:08, Robert Hyatt wrote:
>>>>
>>>>>On September 19, 2004 at 15:10:54, Stuart Cracraft wrote:
>>>>>
>>>>>>On September 19, 2004 at 14:03:52, martin fierz wrote:
>>>>>>
>>>>>>>[snip]
>>>>>>>
>>>>>>>hi stuart,
>>>>>>>
>>>>>>>you seem to be very active programming your engine if the number of posts here
>>>>>>>at CCC is any indication. i guess you have put in all the basic stuff in your
>>>>>>>engine, and now it's playing some decent chess but is getting bashed on by the
>>>>>>>stronger engines, and you're not happy.
>>>>>>>
>>>>>>>i made a similar experience. after years of checkers programming, i started
>>>>>>>writing a chess program, which played it's first game about one month after i
>>>>>>>started on it, and improved in leaps and bounds for a couple of months, which
>>>>>>>was no big surprise to me since i'm rather familar with most of the techniques
>>>>>>>for chess programming from my checkers program.
>>>>>>>after half a year progress got slower, and after 9 months i couldn't detect any
>>>>>>>progress any more. i let it lie around for half a year now, and will try to find
>>>>>>>areas of improvement. but i think that from now on, progress is not so simple
>>>>>>>any more. i will have to take a good look at my evaluation for instance, and at
>>>>>>>the various extensions, and do a lot of testing. i don't believe there is any
>>>>>>>magic bullet that will do the trick. once you reach a certain level, every
>>>>>>>further improvement will have to be earned the hard way.
>>>>>>>
>>>>>>>in this sense, i encourage you to keep working on your engine, without resorting
>>>>>>>to this kind of 'senior programmer to the rescue' attitude!
>>>>>>>
>>>>>>>cheers
>>>>>>>  martin
>>>>>>
>>>>>>I pulled back on an extension idea today and reduced another extension
>>>>>>and the program reached 250 out of 300 on WAC. So that's a "good weekend"
>>>>>>for me at this point. That I have two more days beyond today off to work
>>>>>>on it makes it only taste sweeter. I saved the version under "4.06"
>>>>>>and it is filed in the dusty history book area.
>>>>>
>>>>>
>>>>>You need to think about "cause" and "effect".  IE does doing well on WAC make a
>>>>>program good, or does being good make a program do well on WAC?
>>>>>
>>>>>I believe the latter.  I don't tune for test suites at all.  As the program gets
>>>>>better, it will do better at wac as a result...  Don't be confused and try to
>>>>>make it the other way around...
>>>>
>>>>I believe both.
>>>>
>>>>There are cases that you can be almost sure of improvement in games based on the
>>>>result of test suites.
>>>
>>>"in some cases".  But that isn't quite what is being discussed here.  It is
>>>always "N is better than M if N > M" and that isn't always true.  IE you will do
>>>better at WAC with Crafty, by setting the one-reply and recapture extensions
>>>higher.  But it will play worse as it will extend too much and drop a ply or
>>>more in normal positions...
>>
>>Yes but I believe that you can probably learn from it and extend more single
>>reply in the right part of the cases and not extend in another part of the cases
>>so you can score better both in test suites and not have the problem of droping
>>one ply or more in normal positions.
>>
>>I do not extend single reply the same in all cases.
>>
>>Uri
>
>
>Neither do I.  The deeper into the tree, the less they get extended.

In my case I use the evaluation to decide how much to extend.

The idea is basically not to extend when you are relatively sure about the
outcome of the game.

I think that I can improve it.
I evaluate every node but I do not have function to evaluate different
components and I think that one of the possible improvement that I probably can
get is by remembering a vector of the different components(Gothmog does it) and
using it for better decisions of pruning or extensions and I probably should use
the king safety evaluation to decide how much to extend checks and single reply
to checks.

Uri



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.