An architecture of interval Elman network and its numerical analysis

This paper presents an architecture of interval Elman neural network (IENN) with interval-valued parameters, which can be used to modeling uncertain dynamic systems. The self-feedback links of the context units of IENN provide a dynamic trace of the gradients in the parameter space and enable the network to handle the dynamic modeling of high-order systems. A learning algorithm of IENN is derived in the same way as the error back propagation (BP) method to minimize the cost function. In order to evaluate the performance of IENN, two numerical datasets in different levels of complexity are selected as the modeling targets, and the comparative experiments are conducted with the conventional interval feed-forward BP neural network (IBPNN). The simulation results show that the proposed IENN has better property than the IBPNN in the aspect of approximating.