MODELLING EXPERT JUDGEMENT INTO A BAYESIAN BELIEF NETWORK. A METHOD FOR CONSISTENT AND ROBUST DETERMINATION OF CONDITIONAL PROBABILITY TABLES

Abstract: In project RASTEP (RApid Source TErm Prediction) a computerized tool for real time prediction of source terms at a nuclear power plant is developed. The tool consists of two modules where one is a Bayesian belief network (BBN). A BBN consists of connected nodes and each node has a defined conditional probability table (CPT), which contains the probabilities that a node is in its different states given the states of the node's immediate predecessors. Due to the lack of data the CPTs for some nodes are subjectively determined by experts in the field. Expert judgment may induce uncertainties in the network and it is desirable to know how a relevant and defendable set of conditional probabilities in a BBN can be defined. This Master Thesis is part of a R&D project run by Scandpower on behalf of the Nordic Nuclear Safety Research (NKS). The aim of the thesis is to develop a general method where experts' beliefs can be included in a systematic way when defining the CPTs in the BBN. The proposed method consists of four parts; Network structure, Probability estimation, Sensitivity analysis and Verification and validation. These parts are performed iteratively until the network is robust and reliable. The main focus of the thesis is on the two parts Probability estimation and Sensitivity analysis. From literature different elicitation methods to help the experts assess probabilities in a CPT were found. Two types of elicitation methods were studied; elicitation of a single probability and elicitation of a full CPT. The method preferred when eliciting a single probability was Probability scale since it is an easy and straightforward method for the expert to use. Understanding and implementing methods for generating full CPTs required more attention and were tested both on nodes for example networks and for a network developed in RASTEP. The Likelihood method showed the best result for elicitation of a full CPT and this method is beneficial to use when the expert is uncomfortable at expressing his beliefs as probabilities. An important outcome of the work performed, is that rough estimates of the probabilities are sufficient as a first assignment since a sensitivity analysis will reveal which probabilities have significant effect on the network's output and thus need to be more accurately assessed. The sensitivity analysis also shows the constructor of the BBN how observable nodes, given evidence, influence the network and may lead to modifications in the network's structure.