Optimizing ANN structures for nonlinear system modeling in control engineering

Because of the high complexity and the apparent shortcomings of the customary approaches, nonlinear systems modeling in contemporary control engineering becomes a major challenge. Artificial neural networks (ANNs) are a strong nonlinear approximation method that can be used effectively to address this problem, but how to choose the structure of the artificial neural network and optimize it is a key factor that limits the performance of artificial neural network. In this paper, a hybrid genetic algorithm and particle swarm optimization (GA-PSO) approach of adaptive ANNs structure optimization has been proposed. The approach allows specifying the number of hidden nodes, the kind of the activation functions, and the topology of networks in a dynamic way, ensuring accurate modeling of nonlinear complex systems. An optimization space is built which includes the depth, width, weights of connections and the activation functions in a multi dimension. A better fitness function is used to determine the performance and direct the process of structural evolution on the network. The tests were carried on three common nonlinear control systems, which are the inverted pendulum, the robotic manipulator and the chemical reactor. It has been found that the optimized ANN architectures enhanced modeling accuracy by 23.5%, convergence rate by 35.8%, and generalisation ability by 18.2% over the other traditional fixed-structure neural networks. This paper has given a fresh theoretical underpinning and operational procedure on intelligent modelling of complex nonlinear systems within the discipline of control.