The starch-based film properties database was created with 8 variables and 322 observations collected from the literature. The selected variables were: (1) the starch origin (potato, cassava (tapioca), corn (maize), wheat, yam), (2) the starch concentration, (3) the amylose content, (4) the glycerol concentration, (5) the ambient relative humidity during storage, (6) the aging time of films and two mechanical properties of the starch films at break, (7) tensile strength at break (s(b)) and (8) strain at break (e(b)). The main objective of this work was to classify the data set and to predict mechanical properties (tensile strength (s(b)) and strain at break (e(b)) of starch-based films using a Rival Penalized Competitive Algorithm to find the clusters and, for each class, an artificial neural network (ANN) model from 6 parameters (starch origin, starch concentration (%), amylose content (%), glycerol content, ambient relative humidity (RH) and the aging of films). Each ANN was optimized using a genetic algorithm. The root-mean square error (RMSE) and the coefficient of determination B allowed to choose the best ANN. The results showed that it was possible to distinguish five classes where the composition of each class C-i could be described accurately and connected with the mechanical behavior of the films. This work also showed that it was useful firstly to classify the database before attempting to predict the mechanical properties of the starch-based films. (C) 2013 Elsevier Ltd. All rights reserved.