Reasoning with uncertainty about system behaviour : making printing systems adaptive

Any conclusion about a system’s hidden behaviour based on the observation of findings emanating from this behaviour is inherently uncertain. If one wishes to take action only when these conclusions give rise to effects within guaranteed bounds of uncertainty, this uncertainty needs to be represented explicitly. Bayesian networks offer a probabilistic, model-based method for representing and reasoning with this uncertainty. This chapter seeks answers to the question in what way Bayesian networks can be best developed in an industrial setting. Two different approaches to developing Bayesian networks for industrial applications have therefore been investigated: (1) the approach where Bayesian networks are built from scratch; (2) the approach where existing linear dynamic system theory equations are used as a starting point. A comparison with the traditional method of system identification is also included. The main lesson learned from this research is that Bayesian networks provide a very suitable method for reasoning with uncertainty in engineering problems when detailed engineering knowledge is lacking. When detailed knowledge is available, however, Bayesian networks can still be chosen as an alternative to standard engineering methods. They will offer the advantage that the uncertainty about a problem and its solutions becomes explicit. It will depend on the problem at hand whether their performance will be superior to that of standard engineering methods