Computational cost improvement of neural network models in black box nonlinear system identification

Models play an important role in many engineering fields. Therefore, the goal in system identification is to find the good balance between the accuracy, complexity and computational cost of such identification models. In a previous work (Romero-Ugalde et al., 2013 [1]), we focused on the topic of providing balanced accuracy/complexity models by proposing a dedicated neural network design and a model complexity reduction approach. In this paper, we focus on the reduction of the computational cost required to achieve these balanced models. More precisely, the improvement of the preceding method presented here leads to a significantly computational cost reduction of the neural network training phase. Even if this reduction is achieved by a convenient choice of the activation functions and the initial conditions of the synaptic weights, the proposed architecture leads to a wide range of models among the most encountered in the literature assuring the interest of such a method. To validate the proposed approach, two different systems are identified. The first one corresponds to the unavoidable Wiener–Hammerstein system proposed in SYSID2009 as a benchmark. The second system is a flexible robot arm. Results show the interest of the proposed reduction methods.