Incremental classifier fusion for smart societies

We present an abstract approach to incremental knowledge fusion (classifier fusion) with three different local update rules applied when agents meet. These are: a rule based on the averaging of local information, experience based reputation and transitive reputation, respectively. We introduce and discuss the role of Well Informed Agents (WIAs) in these systems. We analyze each rule in detail and present a comparison that reveals important differences. In particular, best convergence (but with a medium error term) is achieved by the transitive method, whereas middle values of convergence with the smallest error terms are shown by the averaging method. Experience based reputation fares worse of the three, both in terms of convergence speed and error. We discuss consequences for smart societies and directions of future work.