Parameter elicitation in probabilistic graphical models for modelling multi-scale food complex systems

Faced with the fragmented and heterogeneous character of knowledge regarding complex food systems, we have developed a practical methodology, in the framework of the dynamic Bayesian networks associated with Dirichlet distributions, able to incrementally build and update model parameters each time new information is available whatever its source and format. From a given network structure, the method consists in using a priori Dirichlet distributions that may be assessed from literature, empirical observations, experts opinions, existing models, etc. Next, they are successively updated by using Bayesian inference and the expected a posteriori each time new or additional information is available and can be formulated into a frequentist form. This method also enables to take (1) uncertainties pertaining to the system; (2) the confidence level on the different sources of information into account. The aim is to be able to enrich the model each time a new piece of information is available whatever its source and format in order to improve the representation and thus provide a better understanding of systems. We have illustrated the feasibility and practical using of our approach in a real case namely the modelling of the Camembert-type cheese ripening.

► The unified framework of dynamic Bayesian network to model complex system. ► Using of Dirichlet distributions to enrich model whatever the nature of knowledge. ► Taking into account uncertainty and confidence level regarding information. ► Feasibility and the practical use of our hybrid parameter learning in food science.