Best arm identification in rare events