Alam M., Shaheed M., Tokhi M.

This paper presents an investigation into structure optimisation and learning of neural networks with evolutionary algorithms for modelling nonlinear flexible systems. The performance of multi-layer perceptron neural networks (MLP-NNs) largely depends on suitable selection of nodes at different layers, interconnection between nodes and the training algorithm. The first two parameters are often selected heuristically while (commonly) a gradient based algorithm is used to train the network to minimise the error between the target and the predicted output. Gradient based algorithms often suffer from getting trapped in local minima leading to suboptimal solution for the weights and bias values. Evolutionary algorithms with their global optimisation techniques can overcome these problems. In this work, MLP-NN is used to model a single-link flexible manipulator system. This a high order, nonlinear and single-input multi-output (SIMO) system with infinite number of modes each with associated damping ratios, although for practical considerations the first few (dominant) modes are significant. For the manipulator, the input is the motor current while the outputs are the hub angle, hub velocity and end-point acceleration each measured with suitable sensors. Two evolutionary algorithms, namely genetic algorithms (GAs) and particle swarm optimisation (PSO) are used to optimise the structure and interconnection of an MLP-NN trained with backpropagation algorithm. Then GAs and PSO are used to train the optimised MLP-NN. Because of the SIMO nature of the system, the mean-squared error between the three measured outputs and corresponding predicted outputs of the neuro-model are combined to form a single objective for the optimisation process. One-step-ahead prediction of the neuro-model validates the effectiveness of the technique in time domain whereas the frequency domain validations indicate that the model can capture the dominant modes of the actual system well.