Production of high strenght steel alloys is of big importance for the modern methalurgy. The main aim is to obtain high quality materials reducing quantitiy of used expensive compounds. Thus the production of steel alloys starts with optimization of number and content of used alloying components that improve their quality at reasonable price. Since the dependences between input and output variables in that case are strongly nonlinear, application of nonlinear modelling and optimization techniques must be applied.
The fact that neural networks are universal approximations of complex nonlinear dependences that apply “black-box” modeling approach [1, 3, 12] is well known. Therefore, they are proper candidates for modeling structure of such MIMO models. Another useful characteristic of neural networks are their training procedures that are in fact optimization algorithms aimed at minimization of error at neural network output with respect to network connection weights [12, 14]. The well-known backpropagation algorithm is procedure for propagating of derivatives of given function of network output backwards to the input [14]. Thus neural network training