Evolutionary Algorithms (EAs) are search methods inspired by the darwinian theory of evolution, working iteratively on a population of potential solutions that are randomly selected and modified. The selection of a representation, the definition of parameters or the attribution of their proper values have a crucial influence on the algorithm performances. A choice that does not match to the fitness function can make the problem more difficult to resolve. Finding suitable parameter settings is therefore a big challenge. Although EAs are recognized as competitive methods on large problems, they are subjects to certain critics such as parameters adjustment/control. By parameter settings, we mean the approach which consists in finding reasonable parameter values before the algorithm execution. In this thesis, we provide arguments that a set of constant parameters during the run seems to be inadequate. Our contribution to the broad area of optimization concerns the automatic adjustment of parameters according to the test problem. In the first part, we expose the problematic of parameters adjustment/control as well as the principal exisitng heuristics. In the second, we introduce two methods for dynamic control of parameters associated with the representation of solutions. In the third, we propose the States based Evolutionary Algorithm (SEA), a parallel variant of AEs ; this new approach manages simultaneously several EAs in order to control dynamically the parameters during optimization process. In the last part, we present an instantiation of the SEA which integrates different mutation rates in order to adapt the best rate to the search. This new instance was tested on the multidimensional knapsack problem. Comparable results were obtained, which proves that the SEA is capable of dynamically controlling the compromise exploration/exploitation.
