:
Artificial neural networks were used to infer missing data when given geological data were insufficient. But the learning of the traditional networks were too slow for practical use. Authors adopted Fahlman's quickprop program to accelerated artificial neural network, to improve learning ability and speed up the learning time. Some test result showed learning time of new networks was reduced from tens of hours to a few minutes, while the output pattern was almost the same with other studies when the program ran on IBM compatible PC486 DX 66 MHz machine. Cecil's database was used for learning pattern. All the test run showed sufficient accuracy.
Accurate and resonable rock mass classification is needed to design and construct underground rock structures safely and economically. Rock classification methods are empirical ones to classify the rock mass and determine the support amounts by the ratings. Many items are needed like RQD, weathering condition of discontinuities, to classify the rock mass properly. Geological data sets can be usually insufficient in early stage of design because there are many limitations to reach underground site or uncertainty of the property itself. Traditional classification algorithms cannot work when one or more data items are omitted. Even all the data item is given, there can be a confliction between the data. In this case engineer must do some guesswork and inevitably be biased. Artificial intelligence like fuzzy system, expert system or artificial neural networks can be used to diminish the subjective errors. Neural networks are simplified models of neurons. They are composed of nodes in input, hidden and output layers and their connections. They learn facts from the given database through activation function and by controlling connection weights to make generalized inner information. They infer wanted data from new data set which they didn't learn or omitted or from distorted data set which they learned. There are more than 50 neural network models and many of them adopted error back propagation (EBP) learning algorithm with the generalized delta rule by Rumelhart et al.(1986) Some neural networks have been introduced in rock engineering field. Among them are networks to predict elastic compressibility of sandstone specimens (Zahng et al, 1991) and to infer failure mode of rock mass (Lee and Sterling, 1992). Conventional artificial neural networks require exceedingly long learning time. This disadvantage lead to lack of efficiency. They are too slow for practical use. In this study, authors take up new algorithms of accelerated artificial neural network, to improve learning speed and ability of artificial intelligent program. Adopted algorithm is Fahlman's quickprop program (Fahlman, 1989). It is modified and made to a module of ROMES, Rock Mechanics Expert System (Yang et al., 1995). New program was tested by well known XOR and character recognition problems to verify the validity. It was applied to the reference data of Cecil and others which were applied by Lee and Sterling, and Moon and Lee (1993). The inferred results, learning speed and ability were compared with those others.
