Use of Domain Information to Improve the Performance of an Evolutionary Algorithm


In this thesis we explore the use of domain information incorporated during the execution of an evolutionary algorithm, through the use of a cultural algorithm. The cultural algorithms are evolutionary algorithms that support an additional mechanism for information extraction during the execution of the algorithm, avoiding the need to encode the information a priori. Firstly, a cultural algorithm to tackle constrained optimization problems was developed. Such algorithm adopts differential evolution as its model for the population. Using the differential evolution operators as a base, we designed four knowledge sources, each one with a particular influence over the operators. Since each knowledge source exhibits different benefits in different phases of the search, a main mechanism to control the application rate of the operators was developed, based on the success rate each one. This algorithm was tested using a well-known benchmark and a pair of instances of engineering optimization problems, and compared with other representative algorithms of the state-of-the-art. In both cases, equal or better solutions were obtained, requiring a smaller number of objective function evaluations. In the next phase, a hybrid algorithm to tackle multiobjective optimization problems was developed. Such algorithm is a hybrid between the previous algorithm for constrained optimization, and a method of mathematical programming called "-constraint. We obtained other advantages with this algorithm, like obtaining good approaches of the Pareto front in problems that have been very difficult to solve for other evolutionary approaches. As a last contribution, we introduced an approach to perform incor- poration of preferences to the previous algorithm, but such approach can also be used in an wide set of techniques. This proposal is based on the use of vectors of goals. With the addition this approach, is possible to reduce the computational cost needed when applying the hybrid algorithm on problems with a large number of objectives, turning it applicable on practice.