6 -discussion

The algorithms discussed in this paper may be used as a construction set to model the stochastic relation of variables. If for a subset of variables there is a domain theory on their causal dependency this should be formulated as probabilistic rules. For other subsets of variables there may be only an unspeciied association that can be modelled by a neural network with hidden variables. The corresponding likelihood functions are readily obtained and can be combined using the reliability of information as weighting criterion. Appropriate learning algorithms were derived which allow to integrate diierent, possibly connicting, pieces of information. The network can be evaluated by a stochastic simulation procedures.