| Abstract | This paper focuses on the Noisy Iterated Prisoner’sDilemma, a version of the Iterated Prisoner’s Dilemma
(IPD) in which there is a nonzero probability that a “co-
operate” action will accidentally be changed into a “defect”
action and vice versa. Tit-For-Tat and other strategies that
do quite well in the ordinary (non-noisy) IPD can do quite
badly in the Noisy IPD.
This paper presents a technique called symbolic noise de-
tection, for detecting whether anomalies in player’s behavior
are deliberate or accidental. The key idea is to construct a
model of the other agent’s behavior, and watch for any de-
viation from this model. If the other agent’s next action is
inconsistent with this model, the inconsistency can be due
either to noise or to a genuine change in their behavior; and
we can often distinguish between two cases by waiting to see
whether this inconsistency persists in next few moves.
We entered several different versions of our strategy in
the 2005 Iterated Prisoner’s Dilemma competition, in Cat-
egory 2 (noisy environments). Out of the 165 contestants
in this category, our programs consistently ranked among
top ten. The best of our programs ranked third, and it was
beaten only by two “master-slave strategy” programs that
each had a large number of “slave” programs feeding points
to them.
|
