Genetically Optimized Artificial Neural Networks (ANN) Nonlinear System Identification

摘要

Identification is the process of modelling a system based on its inputs and outputs. Identification techniques for nonlinear systems are based on linear approximations of the system, and such approximations perform well for a large variety of processes. But complex systems need sophisticated identification techniques. Neural networks have been shown to outperform traditional identification techniques on complex problems. Neural networks have unique pattern recognition characteristics which enable them to identify nonlinear systems. Genetic Algorithms (GA) have recently been applied to the design of neural networks. Based on the principles of natural evolution, GA leads a more directed search than a random procedure, while still exploring the entire search space. Paper describes technique for optimizing Artificial Neural Networks (ANN) using GA for the identification and control of nonlinear systems. Inverted pendulum (cart-pole) problem is used as the benchmark for this study. Feed-forward and recurrent neural networks are used to model the inverted pendulum. Traditional linear methods, neural networks and neural networks optimized using GA are applied to cart-pole system. And it can be safely concluded that training and optimizing ANNs using GA yield substantially robust designs.