Abstract

Elicitating knowledge from experts is always a dińcult task. Many interviewing techniques exist (N. J. Cooke, 1994), but it is often dińcult to select the right technique for the task at hand. Knowledge elicitation is especially dificult for expert systems that are based on probability theory. The elicitation of probabilities for a probabilistic model of a problem requires a lot of time and interaction between the knowledge engineer and the expert. Through games and other techniques the expert has to be calibrated to get good probability estimates (R. Cooke, 1991).
Bayesian networks (BNs) are an example of a structure that can be used to create a probabilistic model. They consist out of two parts: a graph representing the variables of the model and their conditional dependencies, and conditional probability tables (CPTs) for every node that represent the probabilistic behavior of a variable of the model. BNs need specific conditional probabilities for their CPTs. If an expert does not know these probabilities, but knows other useful probabilistic information, this information generally cannot be used directly to fill the CPTs of the Bayesian network. It will be necessary to perform calculations before the information is transformed into conditional probabilities directly usable for BNs.
Druzdzel and van der Gaag (1995) have proposed a theoretical framework that would allow for the direct use of other types of probabilistic information. This framework has been used as a starting point for the implementation of a non-invasive elication method. Here, non-invasive stands for the ability of the method to directly use any information the expert is willing to state to acquire conditional probabilities for a Bayesian network.
There were many possibilities to improve the framework. Among them were: Finding methods for conflict detection and conflict resolution and improving the sampling process that is at the core of the method. This thesis describes the work and research that was done to implement and improve the method. It describes the performed research, the design and implementation of the method, and an empirical evaluation of the method. The implemented method works, but will need further development to achieve better results.