The introduction of this memory presents briefly artificial intelligence from a personal point of view, insisting on knowledge representation and uncertainty management. We also review some usual human mistakes and biases in the judgement of uncertainty, thus trying to justify why we have used a Bayesian network instead of building a standard --i.e., rule-based-- expert system.
Part II studies the different aspects of BNs, analyzing for every one of them the theory, algorithms and implementation, with special interest in distributed implementations.
With regard to evidence propagation, this memory introduces the graded OR gate, an extension of the binary OR gate which simplifies knowledge acquisition and saves computational time during the inference, and local conditioning, a new algorithm for computing probability in networks with loops, which is much more efficient than previous conditioning methods and in some cases surpasses clustering algorithms.
We also propose a sequential learning model (parameter adjustment) based on Gaussian distributions for conditional probabilities, and a method for the explanation of reasoning, which takes into account two types of interaction (general case and OR gate) and six kinds of links. Part II concludes with a valuation of BNs: history, context, advantages and shortcomings.
Part III describes the construction of the DIAVAL expert system, according to the five classical stages: identification, conceptualization, formalization, implementation and evaluation. We have tried to study in detail the process of knowledge acquisition for Bayesian expert systems in medicine. In the implementation, we have strived to endow DIAVAL with a flexible and user-friendly interface, knowing the importance of this factor for the acceptance of the system by doctors. The evaluation of the the system was satisfactory, but it is still necessary to improve the program before it can be extensively used at a hospital.
Finally, the memory summarizes the lessons learned and the main contributions, assessing the achievements and shortcomings of DIAVAL and showing the research lines open for future work.