Identification of dynamical systems using recurrent neural networks with structure optimization utilizing genetic algorithms

Abstract EN

This paper presents a genetic learning for training recurrent neural networks (RNNs) using series-parallel modeling scheme.

All weights of these networks are adjusted simultaneously with the optimization of RNNs structure.

The evolutionary technique is based on genetic algorithms (GAs) with real coding operators used as a mean for training the RNNs of variable structure (GAs are used for selecting an optimal number of hidden nodes for neuro-identifier, as well as training the network to minimize the error).

The mean square error (MSE) function between thxrpluntand the model output are optimized globally through generations of the genetic search with elitism and hybrid selection method.

Due to the mechanism of a hybrid selection method, elitism strategy and real-coding operators, the GA can find the best model for a given plant early from the first generations.

The significance of the proposed identification approach is illustrated with simulated different examples for linear and non-linear plants off-line.