Optimization of stochastic models of artificial neural networks with Levenberg Marquardt algorithm for the prediction of the concentrations of heavy metals

Abstract - Compared to traditional statistical methods (linear regression, logistic regression, segmentation, discriminant analysis, ...), artificial neural networks are often credited with many benefits such as a better predictive capacity after better representation of the phenomenon (more variables, nonlinear relationships) , adaptability and generalization beyond the studied sample, and also better stability coefficients less sensitive to outliers and the absence of assumptions about the distribution of variables (normalcy regression for example) and yet respecting constraints on the dependent variable. The prediction of levels of Iron, Zinc and Manganese from environmental data in the deposits of the six pits of the Red Sea was performed using multiple linear regression and artificial neural networks. According to this study, we have shown that the best and most appropriate model to be applied for the prediction of heavy metals (Fe, Mn and Zn) in sediments of the Red Sea is the model based on the multilayer perceptron (MLP) type of artificial neural networks, and utilizes a Levenberg Marquardt algorithm as a learning algorithm, and uses the Tangsig and Purelin functions as transfer functions, respectively in the hidden layer and the output layer, and having an architecture [13-8-1]. The results obtained by the two models are interesting. Through preliminary tests, we opted for the best choice of network architecture realized. Artificial neural networks, more efficient, have shown a significant ability of learning and prediction of contents of heavy metals with a very high correlation coefficient and a very low mean square error. The multiple linear regression results are less significant with equations that correlate least the parameters chosen and confirmed afterwards with low correlation coefficients. This study showed furthermore, that the parameters studied on the database are related with the levels of heavy metals by a non-linear relationship.