In this paper dynamic global models of input-output complex systems are discussed. In particular, a dynamic complex system which consists of two nonlinear discrete time sub-systems is considered. As a global model multilayer neural networks in a dynamic structure are used. The global model is divided into two sub-models according to the complex system. A quality criterion of global model contains coefficients which define the participation submodels in the global model. Main contribution of this work is the influence study on the global model quality of these coefficients. That influence is examined for a learning algorithm based on gradient descent method for complex neural networks.
