Artificial Neural Network Trained By Swarm Intelligence Algorithm And Its Applications

The goal of this dissertation is to explore the applicability of Swarm Intelligence Algorithm, such as Particle Swarm Optimization (PSO) and Quantum-behaved Particle Swarm Optimization, to training Artificial Neural Network, particularly Radial Basis Function Neural Network, and to apply QPSO and RBF to Bio-chemical process control. The problem of training neural network can be reduced to maximize the error function calculated through sample data, and it is generally known as a multimodal optimization problem which the gradient-based search technique may fail to search out the global optimal solution. PSO and QPSO are two population-based evolutionary optimization approaches and may play a significant role in training neural network and generate good performance.Firstly, we formulate the principle of some widely known neural network and discuss the concept of Particle Swarm Optimization and Quantum-behaved Particle Swarm Optimization. In turn, we proposed an approach of using PSO or QPSO to train RBF Neural Network, and test PSO-Trained RBF and QPSO-Trained RBF to two paradigms, function approximation problem and problem of forecasting underground water level. The simulation results show that QPSO-Trained RBF can lead to solution with higher precision and faster convergence speed than PSO-Trained RBF.Secondly, we also test QPSO-Trained RBF and PSO-Trained RBF on system identification problem and chaotic time series forecasting problem. The simulation results show that QPSO can identify the system or forecast the time series more precisely and faster than PSO.Finally, we use QPSO-Trained RBF to bio-chemical process control with Genetic Algorithm (GA) Trained RBF also tested for the purposed of performance comparison. The simulation results show that QPSO could enhance the production of amino-acid and generate better solution than GA.The work in this paper indicate that QPSO is a promising training algorithm for neural network and could generate better performance than other intelligent optimization algorithms such as GA and PSO. It will work well on modeling real word problem by neural network.