Optimizing feed forward Neural Networks for Nonlinear Dynamic System Identification with Adaptive Learning rate and Pruning Algorithm

---

Abstract

For effective training and enhancing any ANN's capacity for generalization, structure optimization is crucial. In this work, an effective a pruning strategy was created to identify nonlinear dynamic systems by optimizing the feed forward neural network's structure. The gradient-based back propagation technique is used to update the FFNN weights. A threshold is set and the weights between the input, hidden and output layers below the threshold limit are eliminated. A novel adaptive learning rate is created by calculating the score index Si after each epoch thereby increasing the speed of the algorithm. A nonlinear benchmark problem is used to demonstrate the effectiveness of the suggested algorithm. A comparative study of the algorithm is also done with simple FFNN with ALR and PFFNN with ALR.

---

Keywords

FFNN, pruning algorithm, gradient BP based ANN