Este trabalho mostra que é possível extrair conhecimento útil de dados puros sobre os estudantes de graduação IFMA, de modo a tentar entender os problemas de evasão do referido instituto. Neste artigo, o conhecimento foi modelado como um classificador capaz de identificar quais alunos são os mais propensos a abandonar o curso. Foram usado três algoritmos: Naive Bayes, Support Vector Machine e J48. Assim, baseados no entendimento do problema é possível tomar medidas na tentativa de reduzir essa evasão, como por exemplo, tentar auxiliar o possível aluno evasor antes que isso aconteça, aumentando assim o número de estudantes que se formam.

The purpose of this paper is to show a local search algorithm mixing features of Hill-Climbing, Clonal Selection and Genetic Algorithms. Hill climbing is considered because only the best solution is used. Clonal Selection because the best solution is cloned. Afterwards, individuals are muted using random mutation or non-uniform mutation of genetic algorithms. Four different ways of producing neighborhood solutions have been used in the mutation operator. In the first one (HR), the number of elements are randomly chosen based on the current generation number and muted using random mutation in a certain domain. In the second one (HNU), the number of elements are randomly chosen and muted using non-uniform mutation. In the third one (HRNU), the number of elements are chosen in the same previous way, however the random mutation is used in the initial generations and non-uniform mutation is applied in the last generations. Finally (HNURT), a random number of the elements are muted based on the current generation number, using non-uniform mutation. The performance of the hybrid algorithms is evaluated by means of six multimodal benchmark functions. The results show that HNU and HNURT have better performance. A comparison between the hybrid algorithms and traditional ones, such as, evolutionary strategies, genetic algorithms, particle swarm optimization and differential evolution is presented, as well.

No abstract

No abstract

Evolutionary Algorithms (EAs) are able to find out solutions in many fields and complex disciplines. Parallel Evolutionary Algorithms (PEAs) solve many kinds of problems, as well; moreover it overcomes problems with run time constraints when the problems being solved are much more complex. Thereby, we can state that PEAs can be efficient and faster than a regular EA. On the other hand, parallel programming brings new substantial problems to the developers that have to deal with synchronization and a different debugging process, increasing the learning curve and the programming efforts. In this context, we developed a web-based software that automatically creates java code for parallel genetic algorithms and parallel evolutionary strategies, reducing the required time to develop this kind of application. Two parallel models were implemented, the island model and master-slave. Further, the design patterns strategy and observer were applied in the generation of the PEAs code in order to increase the software maintainability and legibility. An experiment was conducted in order to show how our software can reduce the time consuming of an EA.