Estudo e implementação de algoritmos de inferência e aprendizado em redes bayesianas

Abstract

Bayesian networks are directed acyclics graphs that represent dependences between variables in probabilistic models. This approach represents a good strategy to solve problems that treat uncertainties. The commonest task that we wish to make in a bayesian network is to determine the several interest probabilities conditioned to some events we observe. These probabilities are not stored directly in the model, and consequently need to be computed by inference algorithms. Another important quastion in this approach is that the manual construction of a bayesian net can be very laborious and expensive process for great applications and in complex domains. Therefore, efforts have been driven to development of learning algorithms that can build bayesian nets directly of a database, instead of the specialists’ human discernment. In this work are explained fundamental aspects necessary for the construction of a system able to perform the inference and learning in bayesian networks, it is introduced the system UFLABayes, that allows analysis of the studied algorithms, and are introduced the analyses of the inference and learning algorithms implemented.