| Sumario: | Probabilistic Graphical Models (PGMs) have been widely praised for their declarative nature and their capability for complex reasoning with uncertainty, but when applied to real-world complex domains, the resulting model is usually large and highly inter-connected. This usually brings two main problems: rst, the construction and maintenance of the model turns into a time-consuming, tedious and error-prone task. And second, the computational cost of inference soars with the number of links in the model. Therefore it seems necessary to come up with tools that will alleviate the issues that arise when dealing with large PGMs. In this Master Thesis we have proposed and implemented methods and techniques to help in the process of creation and maintenance of large PGMs. Besides, we describe the process of modelling the problem of programming Cochlear Implants, i.e. adjusting parameters for their optimal performance with the use of PGMs. The new concepts and algorithms we have developed for this purpose are also presented in this Master Thesis. Even if inspired by the needs arisen throughout the development of this real-world application, they are valid for other domains, such as the tuning of systems with adjustable parameters.
|
|---|
