本文主在發展一架構步驟，能規劃學習數值數據，並使用於推理規劃之中。此項功能是由網路神經加權與模糊系統共組而成。主要的特性在於使用多層次的數據分析與學習，同時減少雜訊的產生。最終結果將產生模糊規則與模糊函數。此項結果不僅指示出輸入輸出的變數特性，同時可針對問題產生IF-THEN的輸入輸出關係。此系統經測試後，使用的計算時間，較相似的人工神經網路快了一倍，同時使用經驗數據，也減少了專家系統中，規則詢查繁復的工作。此架構中，容許問題範圍擴大與規則增加，不須重組網路，訓練。只要相對的增加神經元即可。
This study focuses on the development of an algorithm combining self-learning and reasoning with vague numerical data sets. The algorithm derived from neural network and fuzzy system is formed to capture the empirical information commonly inherented in daily life experiences such as stock markets, various manufacturing practices, and many automated control issues. The important feature of the proposed algorithm based on multiple ”reasoning” phases, is to eliminate the ”noises” in the data sets and to construct a fuzzy map (Fuzzy Association Memory) for rule-based reasoning Convergence of ”training” phases is much faster than the normal back-propagation algorithm used in most Artificial Neural Network calculations. It also avoids the rule-matching procedure in developing inference engine of traditional expert systems. The contribution of this work, called ”memory hysteresis,” mainly resides on the expansion of antecedents with corres ponding expansion of neural nets and fuzzy maps. Therefore, the dependent antecedents can be parallelly included, and insignificant rule nodes are faded out as the network converged.
