[computer-go] Improvement of UCT search algorithm

[Darren Cook]

> pairing in discreet rounds.  I will pair top - down.  The best player
> will get an opponent chosen for him probabilistically by the following
> method:
> 
>    1.  Choose best remaining player as player 1.
>    2.  Select N candidate opponents randomly from available players.
>    3.  Choose the BEST of those N candidates as opponent for player 1.
>    4.  Repeat until all players are paired.

I think this has a undesirable quality: if there are 10 strong programs,
and 90 weak programs (which I imagine is typical), the strongest program
will be playing weak programs 90% of the time, so still hardly ever get
to play other strong programs.

I imagine the ideal is playing similar strength programs perhaps 80% of
the time, and playing outside your strength 20% of the time? Does the
swiss-pairing idea give that?

Darren