Estimating a Sparse Representation of Gaussian Processes Using Global Optimization and the Bayesian Information Criterion

Localization in mobile robotics is an active research area. Statistical tools such as Bayes filters are used for localization. The  mplementation of Gaussian processes in Bayes filters to estimate transition and measurement models were introduced recently. The non-linear and non-parametric nature of Gaussian processes leads to new possibilities in modelling systems. The high model  complexity and computation expense based on the size of the dataset are shortcomings of Gaussian process Bayes filters. This work discusses our approach of a sparsing process of a dataset based on Bayesian information criterion model selection and global optimization. The developed approach combines the idea of avoiding model overfitting and Bayesian optimization to estimate a sparse representation of a Gaussian process. Based on visual odometry data of a mobile robot, the method was evaluated. The  results show the operability of the system and unfold limitations of the current implementation such as random-initialization.