A. Kamkar Rouhani; M. Zakeri
Abstract
In order to obtain more accurate results from application of the method of artificial neural networks, instead of selection of the best network determined by trial and error process, we suitably combine the results of several networks that is called committee machine, to reduce the error, and thus, increasing ...
Read More
In order to obtain more accurate results from application of the method of artificial neural networks, instead of selection of the best network determined by trial and error process, we suitably combine the results of several networks that is called committee machine, to reduce the error, and thus, increasing the accuracy of the output results. In this research, ensemble combination of single artificial neural networks has been used in order to estimate the effective porosity of Kangan gas reservoir rock in South Pars hydrocarbon field. To achieve this goal, well logging data of 4 wells in the area at the depth interval corresponding to Kangan formation were used. Acoustic, density, gamma ray and neutron porosity well log data were assigned as the input of the networks while the effective porosity data were considered as the output of the networks. Back- propagation single neural networks having different structures were trained using regularization method and their results were assessed. Then, the networks with the best results, i.e. contained minimum mean of squares of errors in the test step, were selected for making ensemble combinations. To determine the weighting coefficients of the networks in the linear ensemble combinations, we applied three methods of simple averaging, Hashem’s optimal linear combination and non-analytical optimal linear combination employing genetic algorithm, and their results were compared. The best ensemble combination, in which we had the maximum reduction in mean of squares of errors of the test step compared to the best single neural network, was an optimal linear four-network combination obtained by using genetic algorithm optimization method. This best ensemble combination, compared to the best single neural network, reduced the mean of squares of errors in the training and test steps 3.6% and 11.2%, respectively.