Re: computer-go: Perfect play

[Matthew John Saffell <saffell@xxxxxxxxxxx>]

On Fri, 29 Sep 2000 birk@xxxxxxxxxxxxxxxxx wrote:

> >> RoShamBoGod plays "random".   She knows that random play is 
> >> game-theoretic optimal.
> 
> > I thought RoShamBoGod would be better than any mere mortal player.  According
> > to the tournament results, "random" finished 41st in a field of 64 players
> > with a rating of 1800.  The best player had a rating of 2307.  Shouldn't
> > RoShamBoGod have the highest rating?
> 
> The problem here is that it's not a "full information" game. I agree
> that "random" is the best strategy. The only reason that there
> are ratings higher than "random" is that there are also programs
> in the contest that play "sub-optimal".
> Program with higher ratings (than 1800) take advatage of those
> inferior programs, white "random" CANNOT do that ... it's just random.  
> 
> Christoph
> 

I would claim that there is actually a "flaw" (or quirk if you will - or
actually just a consequence of the fact that ratings systems are not
designed to handle perfect players) in the ratings system. No individual
program in the field can beat RoShamBoGod more than a third of the time on
average (she will always win 1/3 draw 1/3 and lose 1/3 regardless of the
opponent's strategy). The reason some have higher ratings though is due in
the same way to the fact that a nonperfect player will alway require a
bigger handicap against GoDevil than against GoGod even though GoDevil and
GoGod always play to a draw.  If you tossed GoGod and GoDevil into any
conventional ratings system, I bet GoDevil would have the higher rating
assuming they play a significant portion of their games against nonperfect
players.  The GameDevil's apparent stature over the GameGod is
proportional to the weakness of the nonperfect players in the game. 

There are apparently a lot of RoShamBoDevil wannabes that are quite
successfull.

Matt