Re: [computer-go] Re: Sharing Secrets

[Don Dailey <drd@xxxxxxx>]

Hi John,

Doing a statistical study of this is an excellent idea.

It's easy  to build a collision  detector.  Oddly enough,  you can use
bits  of  the zobrist  key  for  collision  detection!  Or  you  could
generate a  separate key or do  it the hard way,  actually compare the
positions directly.  But you actually only  need a few extra bits as a
kind of checksum, the more bits  you use the less often the "collision
detector"  is wrong.   But even  4 bits  makes the  detector accuarate
15/16 of the time.

I may  even build this into  my program.  With  a 64 bit key,  I could
have  a collision  detect  mode where  the  program uses  60 bits  for
hashing  and 4  bits for  collision  detection and  the program  could
produce a pretty good statistic on the number of collisions the 60 bit
key is producing.

For Johns study,  I would recommend using few enough  bits for the key
that you will get a fairly large number of collisions.

- Don



   X-Original-To: computer-go@xxxxxxxxxxxxxxxxx
   Date: Thu, 21 Oct 2004 20:18:44 +0200
   From: John Tromp <John.Tromp@xxxxxxxxxxxxxxxxx>
   X-Accept-Language: en-us, en
   Reply-To: computer-go <computer-go@xxxxxxxxxxxxxxxxx>
   Sender: computer-go-bounces@xxxxxxxxxxxxxxxxx
   X-Scanned-By: MIMEDefang 2.42

   Frank de Groot wrote:
   >> You didn't challenge my claim that custom 60 bits is no better than 
   >> random 64 bits.

   > I can't give hard numbers.
   > I have no idea, frankly.
   > What I have done is asked some top-level chess programmers.
   > They said what I just said (do not use random, ever).

   At the webpage I mentioned before they recommend the Mersenne Twister,
   a "twisted generalised shift feedback register".

   Since you were eager to point out the weakness
   of shift feedback register based generators,
   would you care to bet that your custom values (say, 44 bit)
   have fewer collisions than Mersenne Twister values (at say, 46 bits) ?

   Your statement (do not use random, ever) suggests this would be a favorable bet 
   for you.

   It shouldn't be too hard to setup a test where collision probabilities
   can be meaningfully compared...

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

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