The Relief algorithm and its generalizations form a group of filter-based feature evaluation algorithms that are sensitive to feature interactions. We describe the problem of feature selection and present motivation for the application of Relief and its generalizations. We describe all commonly used generalizations of Relief used in classification. We describe the concept of learned metric functions and describe mass-based dissimilarity as well as other learned metric functions, studied in the context of described algorithms. We conclude the thesis with an empirical evaluation of implemented algorithms and metrics. We use the Bayesian hierarchical correlated t-test and plot cross validation results against different cardinalities of feature subsets. We analyze the limitations and assumptions of our evaluation methodology and present ideas for further research.