THEORETICAL AND PRACTICAL ASPECTS OF DECISION SUPPORT SYSTEMS BASED ON THE PRINCIPLES OF QUERY-BASED DIAGNOSTICS

Diagnosis has been traditionally one of the most successful applications of Bayesian networks. The main bottleneck in applying Bayesian networks to diagnostic problems seems to be model building, which is typically a complex and time consuming task.\ud Query-based diagnostics offers passive, incremental construction of diagnostic models that rest on the interaction between a diagnostician and a computer-based diagnostic system. Every case, passively observed by the system, adds information and, in the long run, leads to construction of a usable model. This approach minimizes knowledge engineering in model building.\ud This dissertation focuses on theoretical and practical aspects of building systems based on the idea of query-based diagnostics. Its main contributions are an investigation of the optimal approach to learning parameters of Bayesian networks from continuous data streams, dealing with structural complexity in building Bayesian networks through removal of the weakest arcs, and a practical evaluation of the idea of query-based diagnostics. One of the main problems of query-based diagnostic systems is dealing with complexity. As data comes\ud in, the models constructed may become too large and too densely connected. I address this problem in two ways. First, I present an empirical comparison of Bayesian network\ud parameter learning algorithms. This study provides the optimal solutions for the system when dealing with continuous data streams. Second, I conduct a series of experiments testing control of the growth of a model by means of removing its weakest arcs. The results show that removing up to 20 percent of the weakest arcs in a network has minimal effect on its classification accuracy, and reduces the amount of memory taken by the clique tree and by this the amount of computation needed to perform inference. An empirical evaluation of query-based diagnostic systems shows that the diagnostic accuracy reaches reasonable levels after merely tens of cases and continues to increase with the number of cases, comparing favorably to state of the art approaches based on learning