Information collected from warning systems monitoring natural threats
can be synthesized into a risk measure to determine the state of nature, and for defining the threshold for issuing (or not) of an early warning. This means that a change in the risk measure can trigger the implementation of countermeasures for reducing the hazard, or for reducing the vulnerability and the consequences. Associated costs in the implementation of either should be added to the risk measure, so that the updated risk can be used as the index defining the corresponding warning level.
This work introduces the case of a pre-reliability analysis of the tsunamigenic rockslide at Åknes, Norway. For this purpose, information gathered from engineers, geologists and other stakeholders are incorporated into a probability template based on inference diagrams, which allow for representing causal dependence between events (represented by nodes), taking place from the threat triggering factors up to the definition of the risk measure. Some of these include key events such as the threats triggering factors,
the threats themselves (rockslide and tsunami) and the effect of the Early Warning System. In addition, it is necessary to define how these events interact, which should be stated through the definition of probability distributions.
Once the information from the key events is gathered, it can be resolved using a directed acyclic graph or network, making a model which is graphically more intuitive to understand. In this way, it is possible to trace the passing of information through the network by making use of the dependencies defined for each probability distribution. This is possible either through probability transformation or by letting a set of random variables be a part of a second level of parameters (hyper-parameters). This means that it is possible to create complex information structures using simple causal representations, which help to enhance inferences about the risk measures based on associations.
After constructing the network, the major challenge is to find optimal thresholds for assigning the warning levels. This thesis introduces Monte Carlo simulation techniques to propagate probability states, through variation of threshold levels, so that optimal risk measures with the lowest expected consequence can be found.
Information from the Åknes site is collected and processed in real
time. Bayesian principles will be introduced creating a Bayesian Network which will allow for updating information in any node in the network, and propagate it back and forth on different directions according to the presence of new evidence at any time. This exercise is a pre-reliability analysis, which can help to build the decision-making process associated to the implementation of an Early Warning System.