Adaptive Neural Architecture Reduction for Robust Modeling of Complex Systems

ABSTRACT This study presents a novel approach for identifying nonlinear dynamical systems by developing a modified Jordan recurrent neural network (MJRNN) model. The proposed MJRNN is an extended version of the standard Jordan recurrent neural network (JRNN) architecture. In addition, an adaptive pruning strategy based on the significance of hidden neurons is introduced for artificial neural network (ANN) models to reduce computational complexity and enhance learning performance. The proposed pruning algorithm removes less significant hidden neurons during the training process, resulting in a more compact network structure. Two simulation examples are presented to evaluate the effectiveness of the proposed model. The performance of the MJRNN model is compared with conventional ANN architectures, including JRNN, DRNN, and FFNN, under both fixed and adaptive neuron configurations. The simulation results demonstrate that the proposed MJRNN model with the adaptive pruning strategy achieves superior performance compared with the other ANN models.