[Computer-go] Experiment?

Nick Wedd nick at maproom.co.uk
Thu Mar 18 12:29:36 PDT 2010

In message <299607.55794.qm at web39801.mail.mud.yahoo.com>, terry mcintyre 
<terrymcintyre at yahoo.com> writes
>From: Olivier Teytaud <teytaud at lri.fr>
>  do you think those are program-specific or a general problem with
>  UCT (or more likely the way simulations are run)?
>  Can you post a few specific cases?
>> Big easy semeais. Very clearly defined, so that they are trivial for
>humans, and very big, so that there are many possible
>>  permutations of liberties. One can clearly understand on this kind
>of situations why UCT-like or MCTS-like algorithms
>> will not solve them.
>Yes, I have observed cases of where an otherwise highly-skilled
>UCT/MCTS program will fail, given large well-defined (to humans)

There was such a semeai, misread by both players, in the last KGS bot 
tournament.  See the diagrams in

>Is there any way to do higher-level analysis of such semeais, in the
>same manner as humans do, in order to correctly solve the problem with
>fewer resources?

Maybe.  But first, you need to be able to identify "large well-defined 
(to humans) semeais".

Nick Wedd    nick at maproom.co.uk

More information about the Computer-go mailing list