This paper describes and illustrates the architecture of computer-base
d Dynamic Risk Management Systems (DRMS) designed to assist real-time
risk management decisions for complex physical systems, for example, e
ngineered systems such as offshore platforms or medical systems such a
s patient treatment in Intensive Care Units. A key characteristic of t
he DRMSs that we describe is that they are hybrid, combining the power
s of Probabilistic Risk Analysis methods and heuristic Artificial Inte
lligence techniques. A control module determines whether the situation
corresponds to a specific rule or regulation, and is clear enough or
urgent enough for an expert system to make an immediate recommendation
without further analysis of the risks involved. Alternatively, if tim
e permits and if the uncertainties justify it, a risk and decision ana
lysis module formulates and evaluates options, including that of gathe
ring further information. This feature is particularly critical since,
most of the time, the physical system is only partially observable, i
.e., the signals observed may not permit unambiguous characterization
of its state. The DRMS structure is also dynamic in that, for a given
time window (e.g., 1 day or 1 hour), it anticipates the physical syste
m's state (and, when appropriate, performs a risk analysis) accounting
for its evolution, its mode of operations, the predicted external loa
ds and problems, and the possible changes in the set of available opti
ons. Therefore, we specifically address the Issue of dynamic informati
on gathering for decisionmaking purposes. The concepts are illustrated
focusing on the risk and decision analysis modules for a particular c
ase of real-time risk management on board offshore oil platforms, name
ly of two types of gas compressor leaks, one progressive and one catas
trophic. We describe briefly the DRMS proof-of-concept produced at Sta
nford, and the prototype (ARMS) that is being constructed by Bureau Ve
ritas (Paris) based on these concepts.