Navegando por Palavras-chave "Rbf"

Agora exibindo 1 - 1 de 1
  • Nenhuma Miniatura disponível
    Item
    Somente Metadadados
    Construção Automática De Funções De Kernel Para Support Vector Machines Por Meio De Evolução Gramatical
    (Universidade Federal de São Paulo (UNIFESP), 2017-02-16) Sousa, Arua De Mello [UNIFESP]; Basgalupp, Marcio Porto [UNIFESP]; Universidade Federal de São Paulo (UNIFESP)
    One of the key aspects in the successful use of kernel methods such as Support Vector Machines (SVMs) is the choice of kernel function. A good function is one that maps the input space into a feature space that allows a good separation of the points that compose the dataset. There are several known kernels in the literature that have been manually developed. Among them are the Polynomial, RBF (Radial Basis Function) and Sigmoid kernels. While using these functions can produce satisfactory results for various applications, success in its use is not guaranteed and depends on the choice of parameters associated with these functions. iii Another way to find the best kernel that fits the dataset is th- rough the automatic generation of the function. With this objective the Grammatical Evolution for automatically Evolving Kernel functi- ons (GEEK) was created. GEEK uses a grammar composed of mathe- matical equations extracted from known kernels. When combined th- rough the Gramatical Evolution, these equations give rise to more complex kernels. The results obtained by GEEK were compared in several datasets against the RBF, Polinomial and Sigmoidal kernels, in order to evalu- ate its effectiveness. The grammar used configuration was able to pre- sent results that were not statistically different from the other kernels. Nevertheless, by analyzing solely unbalanced datasets, the GEEK pre- sented a more representative separation of the classes, where the ker- nels compared were biased to the class with more examples.