Bayesian network that learns conditional probabilities by neural networks

In this paper, we discuss a Bayesian network that learns conditional probabilities by neural networks (BNNN). The real world domain is filled with many kinds of variables, which often have uncertainty factors. For the reasoning about such variables, we propose a Bayesian network using feed forward neural networks. Advantages of this model are learning and universal representation capability for the domain mixed with continuous, discrete and multi-dimensional variables. We evaluate the potential of the BNNN on a real world application. Two experiments on learning conditional probabilities with data, which are sampled from the sensors of our autonomous mobile robot and Japanese meteorological statistical records, are introduced.