Hard optimization stands for a class of problems which solutions cannot be found by an exact method, with a polynomial complexity.Finding the solution in an acceptable time requires compromises about its accuracy.Metaheuristics are high-level algorithms that solve these kind of problems. They are generic and efficient (i.e. they find an acceptable solution according to defined criteria such as time, error, etc.).The first chapter of this thesis is partially dedicated to the state-of-the-art of these issues, especially the study of two families of population based metaheuristics: evolutionnary algorithms and swarm intelligence based algorithms.In order to propose an innovative approach in metaheuristics research field, sensitivity analysis is presented in a second part of this chapter.Sensitivity analysis aims at evaluating arameters influence on a function response. Its study characterises globally a objective function behavior (linearity, non linearity, influence, etc.), over its search space.Including a sensitivity analysis method in a metaheuristic enhances its seach capabilities along most promising dimensions.Two algorithms, binding these two concepts, are proposed in second and third parts.In the first one, ABC-Morris, Morris method is included in artificial bee colony algorithm.This encapsulation is dedicated because of the similarity of their bare bone equations, With the aim of generalizing the approach, a new method is developped and its generic integration is illustrated on two metaheuristics.The efficiency of the two methods is tested on the CEC 2013 conference benchmark. The study contains two steps: an usual performance analysis of the method, on this benchmark, regarding several state-of-the-art algorithms and the comparison with its original version when influences are uneven deactivating a subset of dimensions