An adaptive growing pruning algorithm to optimize dynamic feed forward neural networks for Nonlinear Dynamic System Identification

---

Abstract

Optimizing the structure is very crucial for effective identification and control of any nonlinear system. Optimization leads to a robust and more generalized structure. In this work, an effective adaptive growing-pruning algorithm scheme is proposed to optimize dynamic feed forward structures. The hidden layer of the static FFNN is made dynamic and the weights of the dynamic FFNN are trained using standard Back propagation algorithm. Firstly, the network is grown only when the MSE is found high and increasing. Likewise, unnecessarily neurons are pruned based on low activation variance. The learning rate is made dynamic by calculating them based on variance of each neuron. The performance of the proposed algorithm is tested by use of benchmark Mackey glass series problem. The result shows that proposed algorithm performs better than static FFNN with ALR.

---

Keywords

Optimization, FFNN, growing-pruning algorithm, gradient BP based approach