| Summary: | 碩士 === 國立臺灣大學 === 工業工程學研究所 === 88 === The Back-Propagation Network (BPN) is often used a prediction model. The network''s number of learning cycles has a great impact on the network''s learning and prediction performance. An improper number of learning cycles may cause the network under or over fitting the data.
Traditionally, to determine BPN''s number of learning cycles is to observe the Square Root of Mean Squared Error (RMSE) of the learning examples and the testing examples. This is, however, not sufficient from the statistical perspective.
In this research, decision rules based on non-parametric statistical hypothesis tests are proposed. The fitting residuals of the learning examples are tested by run test to examine its randomness. The sign test is then used to test whether the median of residuals is zero. The rational number of learning cycles is determined by the two non-parametric test statistics together with RMSE used in traditional methods.
The proposed method is verified through 22 sets of simulated Box-Jenkins time series data. Results show that it is more effective to determine a proper number of learning cycles using our proposed method. The predicted values are closer to the underlying model. It is also shown that the results of non-parametric statistical testing are consistent with the resulted RMSE from test examples. This result is especially useful when there are insufficient data for BPN''s learning and testing. All the data examples can be used for learning and the proposed non-parametric testing method can be used simultaneously to examine the network''s learning efficiency and to determine a proper number of learning cycles without the RMSE information of testing examples.
|
|---|
