Re: Advice on evaluation

[David Fotland <fotland@xxxxxxxxxxxxxxx>]

Yes, this is much clearer.  I think you need more
reolution than just "can it be killed", and "is
there a sente move".  For many groups in the middle
game, it is possible to tenuki from an attack several times
without being killed.  But after each tenuki, the subsequent 
attack becomes more severe and costs more points.  So 
perhaps a more interesting evaluation is "how many points
do I lose if I tenuki".  In the worst case, you die and
lose points for the size of the group.  Often there
is some small loss, due to later attacks.  And a thick group
can tenuki with no loss of points.

David

At 04:09 AM 5/19/99 +0200, Joan Pons Semelis wrote:
>I did not express myself clearly. My answer to your 
>question about 2 values "what if is B/W turn" was yes,
>the post was a reasoning of why. 
>I've also found Mark Boon problems about playing
>sente moves to keep the 2 unsettled groups from
>dying, it's a common problem with lookahead with
>inperfect evaluating functions that chess people
>call "horizont effect": the program pushes the problem
>beyond the horizont so that it disapears. But making
>simptons disapear doesn't cure the disease, pragmatical
>as it may be.
>My example of ladders might seem to imply "hard" lookahead
>to decide about groups status, it's not necessarily so, I
>completely agree with you in that it can't be used for group
>strength, at least in general, but I did'nt said so:
>
>>From my previous post:
>>  >With group strength the evaluation to decide if it's instable
>>  >it's more complicated than "1 or 2 liberties", the moves to
>
>What I meant is that I want to know for a group the answers to the
>questions " Can it be killed?" and "Are there sente moves against it ?"
>as well as I do with 2 liberties chains of stones. In this later
>case I can find the answer reading all the necessary moves, in the former
>I need to estimate a plausible answer, as actually reading out the problem
>is not practical and in most cases unnecessary, but the idea is the
>same, only in one case the function that answers yes/no is accurate
>while in the other is estimating. How do you use this information for
>deciding on a move is another story, and here is when the "horizont
>effect" takes place if just minimax is used to decide for a move.
>A goal oriented approach seems better.
>
>I hope my point of view is clearer now.
>
>Joan
>
>PS: That function that guesses sente moves against weak groups?
>    Did you heard about the holy grail ? :-)
>
>From: David Fotland <fotland@xxxxxxxxxxxxxxxxx>
>
>>   No, most group strength evaluations can't be decided by
>>  reading.  Take a look at a typical early middle game position
>>  from your own games.  There are lots of stable groups that
>>  don't have two eyes yet.  But if the opponent is a little
>>  thicker, then you want to add a stone.  Evaluating this
>>  relative thickness is very difficult.
>>    
>>  When there is a fight that can be decided by some tesuji, of 
>>  course you can try to read the result.  But you want the
>>  group strength evaluator to first tell you that here is
>>  a fight.  You don't want to waste time reading when your
>>  stones are already dead.
>>  
>>  David
>
>