本研究提出一個複數型模糊類神經系統 (Complex neuro-fuzzy system, CNFS)和採用以資訊理論 (Information theory)為基礎的特徵選取方法應用於分類問題。特徵選取方面以資訊理論為基礎，透過結合最小冗餘和最大相關的概念尋找最佳的特徵子集合。CNFS分類器的建模過程分成結構學習階段和參數學習階段。結構學習階段採用格狀分割法 (Grid partitioning method)，為CNFS分類器挑選重要的模糊規則。參數學習階段使用粒子群演算法 (Particle swarm optimization, PSO)和遞迴式最小平方估計器 (Recursive least squares estimator, RLSE)分別調整模型的前鑑部參數 (Premise parameters)和後鑑部參數 (Consequent parameters)，稱為PSO-RLSE複合式學習演算法，這方法能使模型在建模過程中迅速收斂，達到快速學習的效果。本研究提出的CNFS分類器結合複數型模糊集合 (Complex fuzzy sets, CFSs)和自適應類神經模糊推理系統的架構(Adaptive neuro-fuzzy inference system, ANFIS)，能增加模型的非線性映射能力和提供更靈活的架構。本研究使用美國加州大學爾灣分校 (University of California-Irvine)的機器學習資料庫中十個來自不同領域的資料集來驗證本研究提出的方法，並與其他分類器比較。實驗結果顯示，本研究提出的方法在不同領域的分類問題有優秀的表現。;We present a complex neuro-fuzzy system (CNFS) as a pattern classifier that utilizes complex fuzzy sets. For feature selection of training samples, we consider the removal of redundant and irrelevant features by which we aspire to improve the predictive accuracy of the classifier. Based on information theory, we employ a well-known feature selection method that combines minimal redundancy and maximal relevance for feature selection. One crucial problem for fuzzy-rule based model construction is that the amount of data is usually large in volume, which would make the consequence part parameters of rule base grow exponentially. A modified grid-partitioning method that can select portioned area of input space if some rule-firing-strength threshold is satisfied is employed to deal with that major problem. For the parameter learning method, the particle swarm optimization algorithm (PSO) and the recursive least-squares estimator (RLSE) are integrated as a hybrid learning method to adjust the free parameters of the CNFS effectively. We conducted experiments using 10 data sets of various fields and made performance comparison with other classifiers. The experimental results demonstrate that our approach can find smaller size feature subset with high classification accuracy.
