多階推理模糊神經網路之結構設計

Abstract

本篇論文提出一種多階推理模糊神經網路, 用以學習多階推理的模糊 規則與歸屬函數. 多階推理模糊神經網路是由單階推理模糊神經網路延 伸, 修改而得, 其架構符合多階模糊推理程序.節點函數由 Lasen's MAX- PRODUCT 定義而來. 學習方法是基於回傳式(backpropagation)學習, 並 經由適當的刪除, 得到精簡的模糊規則庫. 針對此多階推理模糊神經網 路, 以三階段討論網路的學習情況. 1.誤差回傳訓練期(the Error Backpropagation Training, EBP-Training, Phase) 2.歸屬函數刪 減期(the Membership Function Pruning, MF-Pruning, Phase) 3.模糊規則刪減期(the Rule Pruning, R-Pruning, Phase)在誤差回傳訓 練期, 多階推理模糊神經網路以回傳式演算,來學得知識. 在歸屬函數刪 減期與模糊規則刪減期, 刪減多餘的歸屬函數與模糊規則, 以求得精簡的 知識庫. 實驗結果顯示, 經過三階段的學習程序, 可獲得精簡的模糊規 則. In this dissertation, a multi-level fuzzy neural network isproposed for learning the firing strengths of fuzzy rules and thebell-shaped membership functions for the linguistic values of inputand output linguistic variables. The proposed network is extendedfrom single-level fuzzy neural network. Basically, structure of proposed network also corresponds to multi-level inference procedure. The node function is designed from lasen's MAX-PRODUCT inference.The backpropagation learning algorithm is adopted to train the learnable parameters. After the deletion of these rules, a fuzzy rule base with precise knowledge and small size can be obtained.A procedure consisting of three different phase for learning the knowledge of multi-level fuzzy neural network is proposed. The threephase are :(1).the Error Backpropagation Training (EBP-Training) Phase,(2).the Membership Function Pruning(MF-Pruning) Phase, and(3).the Rule-Phase(R- Pruning).The learning algorithms in the training phase train the learnableparameters based on the gradient descent concept of backpropagationlearning algorithm. After the training, the MF- Pruning and R-Pruningphase are performanced to delete redundant membership functions andfuzzy rules. Simulation results show three different phase for learningthe knowledge of multi-level fuzzy neural network, a fuzzy rule base with precise knowledge and small size can be obtained.