Re: Superko detection (was Re: [computer-go] M.Boon-Library)

[Don Dailey <drd@xxxxxxx>]

Chrilly,

> Stupid side-Question: What is SuperKo? 

My understanding is  that SuperKo is exactly like  repetition in chess
except that positions are not allowed to match at all.

Positional  SuperKo  is  a  little  different, in  that  it  considers
identical board configurations  as matching, even if the  side to move
is different.

In chess it  makes all the difference in the world  who's turn to move
it is, very often changing a win into a loss or draw.

Under Positional SuperKo it doesn't matter who's turn to move, you are
not allow to make a move that repeats a previous board configuration.

I have been assured that for determining KO, PSK has little effect over
other kinds of KO.

> I have never really understand the point behind GHI. One can reach the same
> position over 2 different variations/graphs. But what bad playing behaviour
> is caused by GHI?

The bad behavior behind GHI has to do with repetition scores stored in the
HASH table, or possibly scoring based on move depths such as mate scoring
which must be handled carefully.     You know all about this.  

Also, as you well know, a common technique to avoid some problems with
GHI is  to not store  repetition draw scores  in the hash  table.  For
instance some programs have a special value used to represent draws by
repetition so that  they avoid using them in the  hash table.  You can
store all  evaluation values with  even numbers for instance  and then
use odd numbers to represent  draws by repetition, depending upon your
contempt factor.   If your  draw by  rep score is  -1 then  you forbid
storing any position with this score in the hash table.

This doesn't  prevent GHI, but makes  it much less of  a problem.  For
some  reason,  if you  avoid  storing draws  in  the  hash table  it's
extremely rare to trace actual  problems to GHI, even though it's easy
to prove  or construct situations where  GHI could be  a problem.  I'm
not sure a general solution has  ever been found to this problem which
doesn't seriously hurt the beneficial effects of using a hash table.


- Don

_______________________________________________
computer-go mailing list
computer-go@xxxxxxxxxxxxxxxxx
http://www.computer-go.org/mailman/listinfo/computer-go/