Local Bayesian Fusion Realized Via an Agent Based Architecture

In the field of reconnaissance and in many other real world applications, information from different possibly heterogenous information sources has to be fused for obtaining adequate results. We present a local Bayesian approach which is realized via an agent based architecture. In analogy to criminalistic investigators, fusion agents elaborate the posterior Degree of Belief of initial hypotheses by local Bayesian modelling and local Bayesian fusion. Thereby, the usually high computational complexity of the Bayesian methodology gets reduced significantly. 1 Bayesian Fusion Methodology The aim underlying Bayesian fusion is inferring the “true value” z ∈ Z of the not directly observable Properties of Interest (PoI) which is to be elaborated in the given fusion task. For an optimal result, prior knowledge d0 and all other available information contributions d1, . . . , dS have to be used comprehensively. Bayesian theory in general bases upon the admissible interpretation of the nature of probability as Degree of Belief (DoB). Conform with this, all involved quantities are assumed as random and modelled probabilistically by conditional DoB distributions. The DoB of an event communicates the degree of its certainty given the available knowledge. At specifying a DoB, useful information may get lost and artifacts may get incorporated. In this case, the resulting DoB is subjective and the final fusion result may be incomplete or distorted. However, the Maximum Entropy (ME) principle [BSW07, Kap93] delivers an established method for the determination of objective DoBs, i. e. DoBs that incorporate the facts completely and have maximal uncertainty simultaneously. In that case, a lossless fusion can be accomplished via the Bayesian theorem [Win97, Zel88]. Its result is the whole posterior DoB p(z|d, d0) of the PoIs given the prior knowl-