Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Evaluation Autotuning

Author: Frank Phillips

Date: 08:42:42 06/29/04

Go up one level in this thread


On June 28, 2004 at 18:29:43, Anthony Cozzie wrote:

>On June 28, 2004 at 16:37:28, Frank Phillips wrote:
>
>>On June 28, 2004 at 16:22:13, Robert Hyatt wrote:
>>
>>>On June 28, 2004 at 16:09:24, Frank Phillips wrote:
>>>
>>>>On June 28, 2004 at 12:43:10, Robert Hyatt wrote:
>>>>
>>>>>On June 28, 2004 at 12:37:42, Dan Honeycutt wrote:
>>>>>
>>>>>>On June 28, 2004 at 08:54:00, Anthony Cozzie wrote:
>>>>>>
>>>>>>
>>>>>>>settings, and then N games with the new settings.  I am only really interested
>>>>>>>in longer timecontrols: 20 min + on an Athlon 2.0G or so (70 min on P-650, etc),
>>>>>>
>>>>>>
>>>>>>Why long time controls?  I thought you could test evaluation with shorter time
>>>>>>controls, search needed longer (or varied) time controls.  Am I out in left
>>>>>>field?
>>>>>>
>>>>>>Dan H.
>>>>>
>>>>>
>>>>>My personal belief is that longer controls are better.  Short games rely heavily
>>>>>on the search, and leaves a better chance for random luck to influence the
>>>>>outcome.  Deeper searches tend to make fewer tactical mistakes, leaving the
>>>>>outcome to the quality of the evaluation....
>>>>
>>>>Two questions for clarification:
>>>>Does this presuppose diminishing returns?
>>>
>>>Not particularly.  What it presupposes is that one search might be more likely
>>>to make errors on shallow depths than another.  IE my simple q-search vs a more
>>>sophisticated q-search.  While at long time controls my q-search appears to work
>>>just fine...
>>>
>>
>>Yes I was assuming the same program (search).
>>
>>>
>>>>And what quality is the evaluation measuring that is different from the prospect
>>>>of future tactics?
>>>
>>>
>>>future tactics != tactics.  Tactics are dynamic.  Evaluation is static.  But if
>>>you think about it, who would be happy using just their evaluation with no
>>>search, to play games?  Why is that?  Because the search is set up to handle
>>>dynamic things by shuffling pieces, the evaluation does better on positions
>>>where everything is static (quiet)...
>>>
>>
>>Yes, how stupid of me. But again I was wondering about diminishing returns.
>>(Theoretically, I would take an infinitely fast searcher over any evaluation
>>function.).
>
>I'll pit my perfect evaluation against your infinite search any day :)
>
>anthony

Ditto my 32 man endgame table.  Sadly, in reality we are left to struggle with
mix and matching imperfect search, imperfect evaluation and <32 man egtbs.  But
that is the fun !



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.