Evidential reasoning using neural networks

A method for using a neural network to model the learning of evidential reasoning is presented. In the proposed method, the belief function associated with a piece of evidence is represented as a probability density function which can be in a continuous or discrete form. The neurons are arranged as a roof-structured network which accepts the quantized belief functions as inputs. The mutual dependency between two pieces of evidence is used as another input to the network. This framework can resolve the conflicts resulting from either the mutual dependency among many pieces of evidence or the structural dependency due to the evidence combination order. Belief conjunction based on the proposed method is presented, followed by an example demonstrating the advantages of this method