Integrating reference point, Kuhn–Tucker conditions and neural network approach for multi-objective and multi-level programming problems