When making decisions about transactions with others, we need to deal
with uncertainty and determine how much trust to place into a
transaction to minimize the risk associated with a bad decision,
i.e. the probability of occurance and the potential impact. Reputation
systems aim at reducing the uncertainty over the behavior of others,
be it people, services, products (as in recommendation systems) or
nodes in self-organized networks, such as peer-to-peer or mobile
ad-hoc networks. Reputation systems work on the premise of being able
to predict future behavior by looking at past behavior. A good
estimation of future behavior is not easy, the method of prediction
needs to be adaptive and deal with erratic behavior, which is by
definition not predictable. Yet, when there are no direct means of
behavior enforcement, we have to try to get the best estimation we can
to minimize expected risk. Our reputation system approach uses a
combination of Bayesian estimation, integration of second-hand
information, weighting, fading of past behavior, and immune-system
inspired secondary response for repeated bad behavior. We find that
reputation systems, when using information from different sources by
leveraging on second-hand information can substantially improve the
estimation of expected behavior based on the past. Reputation systems,
however, introduce a different problem, that of dishonest reports, and
thus have to find a balance between warding of liars and taking
advantage of other participants' information.