Structure and parameter optimization of neural networks using genetic algorithm

A multilayer neural network (NN) can approximate any nonlinear continuous function at the desired accuracy. Hence, neural network are widely used for applications such as system identification, signal processing and prediction, pattern recognition and control. In the applications, fully connected feed forward neural networks including the fixed hidden layer units are commonly used and it is trained using the back propagation algorithm. Small networks having the less units and partial connection may not provide the good performance owing to its limited information processing power. On the other hand, large networks having the more units and full connection between layers may have unnecessary connections, more calculations time and slow convergence of the training process. Furthermore, if the NN has more parameters, the probability of trapping the local minimum of the performance criteria increases. Therefore, structure and parameter optimization of the NN is an important problem.In this thesis, genetic algorithm (GA) is applied to solve the problem of tuning the structure and parameters of the neural networks. For this purpose, the real coding technique of GA is used for optimization of the feed forward neural network and each chromosome is encoded as a vector including the neural network weights, biases and switches for weights and biases. These networks are called as genetic-neural networks (GNN) and ineffective weights and units of GNN can be eliminated with the switches. System identification performance of the GNN proposed in this thesis is investigated. Results obtained from the identification of nonlinear dynamic systems show that the optimized GNN has good identification performance.Keywords: Artificial Neural Networks, Genetic Algorithms, System Identification, Optimization.