A LOCAL-PATCH BASED MULTI-STAGE ARTIFICIAL-NEURAL-NETWORK TRAINING PROCEDURE AND ITS APPLICATION TO MATERIAL CHARACTERIZATION

作者：Luo Yunhua^*; Shah Arvind

来源：International Journal of Computational Methods, 2007, 4(3): 439-458.

DOI：10.1142/S0219876207001266

• Artificial neural network
• local patch
• initial training
• re-training
• multi-stage
• target point

A local-patch based multi-stage artificial-neural-network (ANN) training procedure is proposed in this paper, to improve the accuracy of an ANN trained by a backpropagation (BP) algorithm and, at the same time, to reduce the overall training time. In the proposed procedure a conventional one-stage training procedure is split into multiple stages: an initial training stage and subsequent re-training stage(s). In the initial stage the training data are so selected that the trained ANN has adequate ability of generalization, that is, if provided with a set of new input, the ANN can predict the right region where the output is located, but the accuracy of the solution is not necessarily high. In the following re-training stage(s), local patches of training data, either selected from an existing data pool or generated by numerical methods such as finite element method, are used to retrain the ANN to improve the accuracy. Several factors that may have significant effects on the proposed procedure were investigated based on function approximation. As an example of application, the procedure was then used to train an ANN with finite element data to characterize material parameters.

• 出版日期2007-9