PhD Thesis of Pierre-Charles DANGAUTHIER (2007)
Bayesian probabilities are an efficient tool for addressing machine learning issues.
However, because such problems are often difficult, trade-offs between accuracy and efficiency must be implemented. Our work presents the Bayesian learning method, its philosophical foundations and several innovative applications.
Firstly, we study two data analysis problem with hidden variables. We propose a method for ranking chess players and a collaborative filtering system for movie recommendations.
The second part of our work deals with model learning, with the selection and the creation of relevant variables for a robotic application.
Keywords : bayesian learning, subjective probabilities, generative models, ranking from pairwise comparisons, chess, collaborative filtering, feature selection, model selection, robotics.