Title: Performance comparison of meta-heuristic algorithms for training artificial neural networks in modelling laser cutting
Authors: Miloš Madić; Danijel Marković; Miroslav Radovanović
Addresses: Faculty of Mechanical Engineering in Niš, University of Niš, A. Medvedeva 14, Niš, Serbia ' Faculty of Mechanical Engineering in Niš, University of Niš, A. Medvedeva 14, Niš, Serbia ' Faculty of Mechanical Engineering in Niš, University of Niš, A. Medvedeva 14, Niš, Serbia
Abstract: The application of artificial neural networks (ANNs) for modelling laser cutting is broad and ever increasing. The practical application of ANNs is mostly dependent on the success of the training process which is a complex task. Considering the disadvantages of backpropagation (BP) such as the convergence to local minima and slow convergence, this paper aims at investigating the possibilities of using novel meta-heuristic algorithms such as improved harmony search algorithm (IHSA) and cuckoo search algorithm (CSA) for training ANNs in modelling laser cutting. The validity and efficiency of the algorithms were verified by comparing the results with ANN model trained with real coded genetic algorithm (RCGA) whichs superiority over BP has been well-documented. Statistical methods of the correlation coefficient and absolute percentage error indicate that the search space exploration capability of the IHSA and CSA are comparable to RCGA. It was shown that all three algorithms could be efficiently used for training of ANNs in modelling laser cutting.
Keywords: artificial neural networks; ANNs; metaheuristics; prediction performance; modelling; laser cutting; improved harmony search algorithm; IHSA; cuckoo search algorithm; CSA; ANN training.
International Journal of Advanced Intelligence Paradigms, 2012 Vol.4 No.3/4, pp.299 - 312
Published online: 23 Aug 2014 *Full-text access for editors Access for subscribers Purchase this article Comment on this article