Linear-quadratic cost function for dynamic system modelling using recurrent neural networks

A new idea to improve the performance of neural networks in modelling is presented in this paper. As the networks obtain their knowledge through learning process, it can be influenced through stronger optimization or more suited cost function to be minimized. In this paper, the implementation of linear-quadratic cost function is proposed. This cost function comprises of quadratic and linear function. By applying the linear function to errors greater than a certain threshold, the network becomes more rigid and less sensitive to measurement outliers and disturbances in the form of impulses or spikes. Simulative experiment using a strongly non-linear dynamic system was conducted and the proposed cost function is proved to be effective and enables the network to cope with measurement data with disturbance impulses.