Jeroen Donkers wrote:

> When you write a RoShamBo program you must keep the following in mind. From
> a mathematical point of view, the best approach is to play randomly (with
> equal probability on the three gestures). In the long term, nobody can beat
> you then. However, this does not help you in a competition like this,
> because we do not play long enough. Therefore,
> you should try to predict the moves of your opponent and react on that, or
> you should try to be smarter than your opponent.

        Even more to the point: in a competition like this, the goal is to do
best, not to play a moderate, reliable game. Thus random play is an
unstable fixed point; if one program plays nonrandomly, some programs
that attempt to capitalize on that will (if competently written) beat
programs that don't. Some of them will in turn be lunch for more cunning
programs, of course. But any program that settles for an easy one win in
two will *not* be in the winners' circle and will gain no bragging
rights for its creator; thus there will be a motivation to write
"aggressive" programs.  

        
        Of course, some limitation on play is necessary, as we cannot declare a
winner in an infinite tournament (imagine each move played in half the
time of the previous one!) unless the limiting proportion of wins is not
1/2; and there is an obvious draw-forcing strategy of resorting
permanently to random play when (say) one has been down a million
points. Nonetheless, the value of random play is perhaps less than one
might assume when one plays to win. (This is why stock-picking contests
are usually won by strategies that you wouldn't try in real life, such
as putting everything on a penny mining stock. It's more likely to
triple its value in a year than, say, IBM.)

        -Robert Dawson
.
.
=================================================================
Instructions for joining and leaving this list, remarks about the
problem of INAPPROPRIATE MESSAGES, and archives are available at:
.                  http://jse.stat.ncsu.edu/                    .
=================================================================

Reply via email to