自適應機率神經網路

Adaptive Probabilistic Neural Networks

本研究提出自適應機率神經網路（A daptive Probabilistic Neural Networks, APNN），它包含三種參數：代表變數重要性的變數權值、代表樣本有效範圍的核寬倒數、及代表樣本可靠程度的資料權值。本研究提出自適應調整這些參數的演算法，藉由學習過程優化這些參數，使模型的準確度最佳化。為證明此網路的性能，本研究以三個人為的函數映射問題以及一個實際的分類問題來做測試，並與倒傳遞網路（BPN）及機率神經網路(PNN)做比較。結果證明自適應機率神經網路的模型準確度只略低於BPN，而遠優於PNN，且APNN的變數權值確實可以顯示輸入變數影響輸出變數的重要程度，使模型具有部份解釋能力。

This study proposes adaptive probabilistic neural networks (APNN), which include three kinds of parameters: the variable weights representing the importance of input variables, the core-width-reciprocal representing the effective range of patterns, and the data weights representing the reliability of patterns. This study proposes a algorithm to adapt these parameters, and maximize the accuracy of the model by optimizing these parameters in the learning process. In order to prove the performance of APNN, three artificial function mapping problems as well as an actual classification problem are employed to test it and compare it with a back-propagation network (BPN) and a probabilistic neural network (PNN). The results proved that the accuracy of APNN is only slightly lower than BPN, and is strongly superior to PNN. Furthermore, the variable weights of APNN really express the importance of input variables for output variables, which provides the model with explanation abilities.