Título: Mean Derivatives Based Neural Euler Integrator For Nonlinear Dynamic Systems Modeling
Autores: Tasinaffo, Paulo M.; Rios Neto, Atair
Resumo: The usual approach to nonlinear dynamic systems neural modeling has been that of training a feed forward neural network to represent a discrete nonlinear input-output NARMA (Nonlinear Auto Regressive Moving Average) type of model. In this paper, the recently developed alternative approach of combining feed forward neural networks with the structure of ordinary differential equations (ODE) numerical integrator algorithms is done in a way not yet considered. In this new approach, instead of using the neural network to learn the instantaneous derivative function of the ordinary differential equation (ODE) that describes the dynamic system, it is used to learn the dynamic system mean derivative function. This allows the use of an Euler structure to obtain a first order ODE neural integrator, which in principle can provide the same accuracy as that of any higher order integrator. The main objective is to have an approach in which the dynamic system neural modeling is simple. First in terms of the feed forward neural network training, since it has to learn only the algebraic and static functions of the system dynamic ODE mean derivatives. Second in terms of numerical complexity, since a first order integrator structure is sufficient to attain a specified accuracy. Test results of a practical problem, representing the dynamics of orbit transfer between the Earth and Mars, are used to illustrate the effectiveness of this new methodology.
Palavras-chave: Dynamic Systems Modeling; Neural Models; Numerical Integrators; Feed Forward Neural Nets
Páginas: 12
Código DOI: 10.21528/lmln-vol3-no2-art5
Artigo em PDF: vol3-no2-art5.pdf
Arquivo BibTex: vol3-no2-art5.bib