Improved neural network modeling approach for engineering applications

The Muti-Layer Perceptron (MLP) neural networks (NN) trained with back propagation algorithm (or BPNN) is the most popular and widely used network paradigm in engineering applications. Two setbacks of NN modeling are (1) very slow rate of learning and (2) trial-and-error problem-dependent selection of architectural/learning parameters. This paper focuses more on the trial and error methodologies used in NN engineering applications. An improved trial and error modeling approach is proposed to synchronize the optimization of BPNN's architecture and learning performance. In addition, the new approach addresses several other common concerns on the use of the BPNN technique in engineering applications, such as (1) detecting and avoiding over-fitted models; (2) decoding the state of intelligence of "black box" NN models; and (3) augmenting the ability of NN to extrapolate beyond the training data. A case study of applying BPNN to characterize the geographical terrain slope is presented to illustrate the advantages of the new modeling approach. A total of 66 patterns were captured from a terrain geographical information system (GIS) recently developed for investigating six test areas in the Losses Plateau, Northwestern China. A computer program was coded in house to implement the proposed approach and arrive at the final model for characterizing the mean slope. The model is able to extrapolate beyond the training data to a certain extent and demonstrates "intelligent" reasoning capability in relating inputs to the output.