Resumo

NEGRIN DIAZ, Iván A.; NEGRIN HERNANDEZ, Luis I.  e  CHAGOYEN MENDEZ, Ernesto. Parameter tuning of genetic algorithms: composite method proposal. Rev cuba cienc informat [online]. 2020, vol.14, n.3, pp.59-82.  Epub 01-Set-2020. ISSN 2227-1899.

Parameter tuning deals with finding the best configuration of an optimization method in a given problem. It is an extremely high computing process. Some problems, such as structural optimization, need enormous resource consumption, so tuning the method´s parameters in these processes is certainly expensive. One way to avoid this drawback is to use analytical functions (or benchmark functions), simulating the main features of real ones. In this paper, the Eggholder function is used as case study to tune the parameters of the Genetic Algorithms (GA), using as utility the graphical visualization of the average performance curve, and its correspondent MBF value. The results showed that for optimizing these high-complexity functions, it is necessary to establish high population sizes (300 or more). The best configuration reached was using uniform selection, heuristic crossover, reproduction establishing an elitism of 15 % of the population size and a crossover fraction of 0.60. Due to the inability of simple GA of finding the global optimum regularly, other solutions are recommended. The first one is using discrete variables in the optimization process. The second one consists on creating an initial population, using the simple GA itself. This is what we call compound method, and it was capable to find the global optimum the 100% of the tests, demanding only 16% more computational consumption.

Palavras-chave : optimization; parameter tuning; Genetic Algorithms; population; selection; crossover.

        · resumo em Espanhol     · texto em Espanhol     · Espanhol ( pdf )