This paper describes the design of a normative engineering risk management system that performs four activities: diagnosis, dynamic evolution, decision making, and information gathering. The problem domain can be represented as a partially-observable Markov decision process (POMDP). By decomposing the POMDP into a separate model for each of the four activities, the system enables highly asymmetric models to be tailored to a particular situation. Our goal is to construct a normative system that creates an accurate, concise decision model and that provides timely, reliable results. Conserved model components are constructed at design time whereas unique features are customized at run time. For run-time model components, human guidance is used sparingly to refine the decision model. Customized asymmetry is most extreme for information gathering. An application to the management of fire risks on board offshore oil platforms, with particular emphasis on gas leaks, illustrates the concepts. (C) 1997 Elsevier Science Limited.
