But how can I compute all those probabilities for the input variables? How can I decide, with what probability the rock of a burned forest is permeable or impermeable?
	Easy! You look at the source of your data and decide how reliable it is. Stassopoulou, Petrou and Kittler in a paper they published in the International Journal of Geographical Information Science, 1998, vol 12, no 1, pp 23-45, discussed in detail how to estimate the uncertainty in each input variable.
	And how can I compute all those conditional probabilities?
	That is the problem!  You need lots and lots of training data to cover all possible combinations of states of conditions and effects. Stassopoulou and Petrou in a paper published in the International Journal of Pattern Recognition and Artificial Intelligence (Click here for an acrobat version of the paper), mapped such a probabilistic network on a neural network, in order to take advantage of the ability of a neural network to be trained.
	So, let me understand: I construct a probabilistic network. I map it on to a neural network, that is I find the correspondence between the conditional probabilities used by the probabilistic network and the weights used by the neural network. Then I train the neural network using the training data that are available, which means I determine the values of the weights of the neural network. Then from the values of the weights I determine the values of the conditional probabilities  of the probabilistic network, and then I use the probabilistic network to make decisions!
	Exactly!!