In the real world, information is sometimes observed incompletely. 
Then, an intelligent system working in the real environment 
has to combine observed information and prior knowledge of the environment.
In order to realize this integration mechanism, a model based on Bayesian 
decision theory is proposed. 
In this framework, prior knowledge is represented as
a conditional prior probability distribution, and integrated by Bayes' theorem.
For optimum Bayesian decision theory based on a posterior probability,
we have to approximate proper prior probabilities, which sometimes depends 
on particular situations. 
These situated prior probabilities can be modeled by Bayesian networks. 
Especially, Bayesian network on neural networks (BNNN) has some advantages 
for learning situated prior in the real world environment. 
As the demonstration of the framework, learning situated prior 
probabilities from an English dictionary is shown.
It can apply to handwriting character 
recognition using Bayesian classifier with a neural network. 
The result improves recognition performance better than an ordinal 
(maximum likelihood) classifier.