Belief fusion: Aggregating pedigreed belief states (Q5946339)

The problem of multi-agent belief aggregation is considered. A new operation, called belief fusion, for combining the beliefs of agents informed by some set of sources which are ordered by reliability is introduced. An (anonymous) belief state is defined as a total pre-order over the set of possible worlds (interpretations) \(\mathcal W\). The belief fusion operator accepts two belief states and produces a third. Moreover, the operator adjudicates between the two belief states where they disagree. As a tool for resolving conflicts a refinement relation on belief states is defined. The next step of the construction consists in an enrichment of belief states by a representation of the source of each information item. A pedigreed belief state induced by a set of sources \(S\) is defined as a function assigning a subset of \(2^{S}\) to a pair of possible worlds. A policy for resolving conflicts within a pedigreed belief state is defined. The fusion of pedigreed belief states \(\Psi_{1}\) and \(\Psi_{2}\) induced by sets of sources \(S_{1}\) and \(S_{2}\) is defined as the pedigreed belief state induced by \(S_{1} \cup S_{2}\). Iterated belief fusion obeys the invariance properties of idempotence, commutativity and associativity. The set of pedigreed belief states that is closed under fusion forms a semilattice. Fusion of two pedigreed belief states returns the least pedigreed belief state that contains at least as much information as both of them. A symmetric operator is defined, which returns the greatest state containing no more information than either one. Finally, the relation between fusion and AGM revision is discussed. AGM revision reduces to the fusion of partially specified belief states of a novice agent and an expert agent. No specific source information is given for either agent, but it is assumed that the expert is better informed. However, when regarding a reduction of fusion to revision, existing iterated revision operators applied as fusion do not satisfy the same invariants that the fusion operator does (a series of propositions of the form ``The belief sets resulting from the left and right association of Boutilier's natural operators can be inconsistent'' is proved).