通用迴歸神經網路在中長期需求模式上之建構與探討-----以電腦零組件為例

Implementation of General Regression Neural Network into Long-Term and Middle-Term Demand Forecasting Models--------A Case Study for Computer Components

邱穎聖

指導教授 ： 陳雲岫 博士

元智大學/工程學院/工業工程與管理學系/碩士(2002年)

摘要

本論文提出以通用迴歸神經網路(General Regression Neural Network)建構一需求模式，並預測電子資訊產品之未來市場需求量。GRNN是從PNN(Probability Neural Network，機率類神經網路)所演變而來，主要應用在預測及控制上，可用來建立連續變數之函數關係，無論迴歸問題為線性或非線性均可用GRNN來解決。單變量時間序列、指數平滑法、迴歸模式是較常被使用在需求預測的方法。本論文則採用GRNN之模式，並與傳統之預測模式及整合模式比較，分別找出各預測模式最佳的參數組合，以探討GRNN是否比其他四種預測方法快且準確性高；本論文針對主機板銷售量進行實證研究，以經濟部統計處出版的工業生產統計初步速報作為研究驗證的研究資料。預測結果顯示，整合預測與GRNN之結果並無顯著差異，但是整合預測在進行學習時，花費較多時間，而GRNN學習及預測速度快，且比其餘的三種模式均來得佳，故GRNN在需求預測上有很好的表現。

關鍵字

GRNN ； 單變量時間序列模式 ； 指數平滑法 ； 迴歸模式 ； 整合模式 ； 需求預測

並列摘要

In this thesis, we implement General Regression Neural Network into a demand model to forecast the demand of long-term and middle-term electronic products. The GRNN is a evolution from Probability Neural Network (PNN) and applied in control and forecasting problems for finding the relationship between continuous variables. One of the merits of GRNN is that GRNN can fit either linear or non-linear regression lines without extra efforts. Conventional statistical methods to forecast demands are unvariate time series, exponential smoothing method and regression analysis. In the research, performance evaluation of GRNN is studied and the comparisons among unvariate time series, exponential smoothing method, regression analysis and a weighted combination of several forecasting models are conducted. Mean absolute percent error (MAPE) and mean absolute deviations (MAD) are two performance indices used in the research. Studies show that both MAPE and MAD of the weighted linear combination models are smaller than any other models except GRNN model. There is no significant difference of MAPE and MAD between GRNN and weighted linear combination models. From time consumption viewpoint, GRNN is much shorter than the weighted linear combination models, thus the proposed GRNN model performs well in demand forecasting.

並列關鍵字

Key Words：General Regression Neural Network ； Unvariate Time Series ； Exponential Smoothing Method ； Regression Analysis ； Weighted Linear Combination Models ； Demand Forecasting

參考文獻

8. 劉亭宜，「GRNN在晶圓製造裡良率模式之建構與分析」，私立元智大學工業工程研究所碩士論文，中壢(2000) 。

9. Disorntetiwat, P. and C. H. Dagli, “Simple Ensemble-averaging Model Based on Reneralized Regression Neural Network in Financial Forecasting Problems,” Adaptive Systems for Signal Processing, Communications and Control Symposium 2000 IEEE AS-SPCC, 2000, pp. 477—480.

10. Dong, J., “Reseach on the Method of Non-linear Combining Forecasting Based on Fuzzy Neural System,” Proceedings. of the 3th World Congress on Intelligent and Automation, 2000, pp. 899-903.

11. Frost, F. and V. Karri, “Performance Comparison of BP and GRNN Models of the Neural Network Paradigm Using a Practical Industrial Application,” Proceedings. of ICONIP 1999 6th International Conference on Neural Information, Vol. 3, 1999, pp. 1069-1074.

12. James, T. L., O. R. Jens and S. Brian, “A Hybrid Econometric-neural Network Modeling Approach for Sales Forecasting,” Tourism Management, Vol. 27, 1996, pp. 175-192.