Keywords: artificial neural networks, PSO, civil engineering deformation, prediction.

Artificial neural networks have been applied in many instances successfully in which conventional mathematical modeling technologies are not accurate or capable, because of its capability of nonlinear analysis. Yet the traditional back propagation (BP) algorithm convergence is very slow, and in turn often requires hundreds and thousands times of iterations. With the increase of the dimension of samples, the speed of convergence of the network will in turn decrease. Routine training algorithms such as BP or other gradient algorithms always result in very slow convergence and easily get stuck in a local minimum. Initial connection weights, learning parameter, inertial weight and network structure are the factors that affect the accuracy of the prediction. To determinate the proper topology is essential for the application of artificial neural networks, especially for calculating the number of hidden nodes. The design of neural networks mainly consisted of two parts: one is topology design, including numbers of hidden layers, numbers of nodes and the ways nodes are connected; the other is the training of neural networks. So, some algorithms are put forward for determining the structure of neural networks. Nowadays, more and more research has been focusing on the study of evolutionary computation that revamps the artificial network.

Differentiating from traditional search algorithms, evolutionary computation techniques work on a large sum of potential solutions for the search space. Through the cooperation and competition among the potential solutions, the applications of these techniques to the complex optimization problems can often identify optima more quickly. As a kind of evolutionary computation, the particle swarm optimization (PSO) algorithm can be used to evolve and calculate three aspects of the artificial neural network: connection weight, the topological structure of network, and the network learning algorithms. The present study in terms of this technique is specified in combing the PSO algorithm and the BP algorithm to train the neural network. The advantage of the PSO over many of the other optimization algorithms is its relative simplicity and rapid convergence.

In this paper, the practice of the PSO algorithm for optimizing neural networks is presented. The paper discusses the advantages of applying the PSO algorithm and the BP algorithm to the train neural network against the sheer application of the BP algorithm in civil engineering displacement prediction. The study focuses on training the connection weights between the neurons in different layers and the number of hidden nodes of the BP neural networks through the PSO algorithm in order to make displacement predictions for a pile in a deep pit. The result indicates that the accuracy and convergence velocity processed by this method is much better than when only the BP algorithm adopted.

purchase the full-text of this paper (price £20)

go to the previous paper
go to the next paper
return to the table of contents
return to the book description
purchase this book (price £80 +P&P)