Premium
Melt index prediction by adaptively aggregated RBF neural networks trained with novel ACO algorithm
Author(s) -
Li Jiubao,
Liu Xinggao,
Jiang Huaqin,
Xiao Yundu
Publication year - 2011
Publication title -
journal of applied polymer science
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.575
H-Index - 166
eISSN - 1097-4628
pISSN - 0021-8995
DOI - 10.1002/app.35688
Subject(s) - overfitting , artificial neural network , computer science , radial basis function , algorithm , hierarchical rbf , ant colony optimization algorithms , artificial intelligence , process (computing) , data mining , mathematical optimization , mathematics , operating system
Three estimation models of polypropylene (PP) process to infer the melt index, an important quality indicator determining product specification, are presented. Radial basis function (RBF) neural network (NN) is used to develop the models because of its capacity of fitting the complex relationship in PP process. A novel ant colony optimization (ACO) algorithm is also proposed and used to solve the optimization problem of the continuous linking weights when training the RBF NN. Based on the RBF NN and the novel ACO algorithm, a single NN model is developed. However, a single network cannot always work well due to some defects (such as overfitting) of a NN. Thus, as an improvement of the single NN model, several RBF NN trained with a certain objective are combined, and the aggregated NN model is obtained. To make the aggregated NN more robust and effective, an adaptive method of assigning the combinational weight to every individual network is applied to the former aggregated NN model and finally an adaptive aggregated NN model is achieved. Further researches of the three models are carried out on the data from a real industrial plant, and the prediction result shows that the performance of the obtained prediction models is better and better with every improvement step taken as above. The adaptive aggregated NN model works best, and the satisfying prediction error it provides depicts its prediction accuracy and universality, as well as an application prospect in PP process. © 2011 Wiley Periodicals, Inc. J Appl Polym Sci, 2012