**Título:** Replicação bootstrap e Análise de sensibilidade em Redes Neurais Artificiais

**Autores:** Almeida, Renan M. V. R. de; Infantosi, Antonio Fernando C.; Gismondi, Ronaldo C.

**Resumo:** The relative importance of the input variables in an Artificial Neural Network (ANN) model was investigated by means of bootstrap replications and a sensitivity analysis. The ANN was developed from 43 variables of the social, economic, environmental and health sector of 59 Brazilian municipalities, using infant mortality as the dependent variable. Eight variables, chosen with the help of a Factor Analysis on the data, were used as inputs. The relative importance of the inputs was investigated with the help of bootstrap replications of the model. The determination coefficient R 2 and the Mean Square Error (MSE) were obtained from bootstrap samples for the entire model, and then compared to those obtained after withdrawing each input variable at a time. For the full model, it was found R^2 =.80 and MSE = (7)(10^3). The sensitivity analysis identified as the most important variables literacy, agricultural sector jobs and number of commercial establishments. The method seems a simple alternative for evaluating input importance in ANN models.

**Palavras-chave:**

**Páginas:** 6

**Código DOI:** 10.21528/CBRN2001-066

**Artigo em pdf:** 5cbrn_066.pdf

**Arquivo BibTex:** 5cbrn_066.bib