Counter-Propagation Artificial Neural Networks (CP-ANNs) require an optimisation step in order to choose the most suitable network architecture. In this paper, a new strategy for the selection of the optimal number of epochs and neurons of CP-ANNs was proposed. This strategy exploited the ability of Genetic Algorithms to optimise network parameters. Since both Genetic Algorithms and CP-ANNs can lead to overfitting, the proposed approach was developed taking into considerable account the validation of the multivariate models. Moreover, a new criterion for calculating the Genetic Algorithm fitness function was introduced. The percentage of correctly assigned samples for calibration and internal validation were both used in the optimisation procedure, in order to get simultaneously predictive and not overfitted models.The optimisation strategy was tested by the use of several chemical benchmark data sets for classification tasks and results were compared with those of the exhaustive searching of all the possible solutions. (c) 2010 Elsevier B.V. All rights reserved.
Ballabio, D., Vasighi, M., Consonni, V., Kompany Zareh, M. (2011). Genetic Algorithms for architecture optimization of Counter-Propagation Artificial Neural Networks. CHEMOMETRICS AND INTELLIGENT LABORATORY SYSTEMS, 105(1), 56-64 [10.1016/j.chemolab.2010.10.010].
Genetic Algorithms for architecture optimization of Counter-Propagation Artificial Neural Networks
BALLABIO, DAVIDE;CONSONNI, VIVIANA;
2011
Abstract
Counter-Propagation Artificial Neural Networks (CP-ANNs) require an optimisation step in order to choose the most suitable network architecture. In this paper, a new strategy for the selection of the optimal number of epochs and neurons of CP-ANNs was proposed. This strategy exploited the ability of Genetic Algorithms to optimise network parameters. Since both Genetic Algorithms and CP-ANNs can lead to overfitting, the proposed approach was developed taking into considerable account the validation of the multivariate models. Moreover, a new criterion for calculating the Genetic Algorithm fitness function was introduced. The percentage of correctly assigned samples for calibration and internal validation were both used in the optimisation procedure, in order to get simultaneously predictive and not overfitted models.The optimisation strategy was tested by the use of several chemical benchmark data sets for classification tasks and results were compared with those of the exhaustive searching of all the possible solutions. (c) 2010 Elsevier B.V. All rights reserved.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.