An Adaptive Resonance Theory Based Intelligent Neural Fuzzy Forecasting System

• Main Authors: Liu, Ying-Jang, 劉瑛展
• Other Authors: Chang, Chir-Ho
• Format: Others
• Language: zh-TW
• Published: 1998
• Online Access: http://ndltd.ncl.edu.tw/handle/71623231678791220534

碩士 === 華梵大學 === 工業管理研究所 === 86 === 　　The purpose of this research is to examine how to use an improved artificial neural network and fuzzy model to forecast the inventory amont/demand based on a historic time series data set. We first examined the existing adaptive resonance theory based networks to see if they are suitable to solve our problem. A continuous data set was processed and subdivided by using a sliding window of fixed-length before analysis is conducted. Taking the advantages of ARTs, it is verified that plasticity and stability are maintained. Unfortunately, when apply existing ART-based neural networks to classify the patterns, we have fair results. These results: partially came from an unproper bit-map coding scheme, or partially came from thelessential confontation of binary/analog data. 　　We proposed a time-space trainning algorithm for an ART-based neural network. The algorithm differs with earlier versions of training in several aspects: 1) the use of cosine angle to depict the similarity of two vectors, 2) no rrecursive iterations as ART2 are needed, and 3) the dynamic data set needs no binarizing. 　　The inventory demand forecast is performed via a combining of artificial neural and fuzzy ststem. Whenever there is a new pattern developed (over a pre-specified vigilance paremeter) during the dynamic learning process, a new rule is automatically established in our fuzzy system and added to the knowldege base where historic data acted as domain expert. Simulation results show that the proposed time space embedded ART nerual model outperforms those use ART1, ART2A, and ART2 conventional algorithms in pattern classification when using a pseudo-inventory data set of one hundred terms. Besides, the intelligent forecast system can predict quite correctly (within a reasonable RMS value) up to a consecutive up to six terms into the unknown future. Since forecasting problems, in general, involve a lot of factors, more time and energy could be dedicated to the multi-ART, multi-pattern, advanced fuzzy systems.