| Summary: | 碩士 === 國立臺灣大學 === 土木工程研究所 === 82 === It have been recognized that the expert systems can solve do-
main-specific problems as good as human experts. However, the
kn- owledge acquisition bottleneck is the most difficult
problem in building expert systems. Some researchers developed
machine lear- ning techniques to reduce the efforts of
knowledge acquisition. The traditional machine learning is
symbolic-oriented. The know- ledge learned by traditional
machine learning is existed in an explict form. Most machine
learning mechanisms are weak whenever the training examples
contain noise. Thus, another research dir- ection set to
artificial neural networks, which use numerical co- mputation
approach to mimic the human neuron. The acquired knowl- edge of
artificial neural networks is existed in an implicit form . The
development of artificial neural networks was declined af- ter
Misky and Papert''s "Perceptron". Minsky proved that the arti-
ficial perceptrons could not solve some simple problems such as
"XOR" problem. However, the XOR problem was solved by the
hidden layer and backpropagation algorithm in 1986. This
research uses the example-based backpropagation algori- thm in
neural networks as a learning machanism. After the layout of an
artificial neural network is generated, the network can be
trained by examples one by one. Then, the network induces some
common features of input examples. The knowledge domain of this
research is prestressed concrete bridge design. Firstly, a
prest- ressed concrete bridge design system is constructed and
it produ- ces results for the design. Thirdly, the output
results are impl- emented as the inputs of the artificial
neural network, and the common design knowledge are derived.
Meaning, the neural network is a tool for reconstructing the
hidden knowledge of the system mentioned. Finally, the
advantagess and restrictions of using ne- ural networks as a
knowledge acquisition paradigm are discussed.
|
|---|
