利用基因演算法根據不同的應用自動設計並建構具有功能的隱藏層節點的類神經網路

Abstract

在傳統的類神經網路 (Artificial Neural Networks, ANNs) 裡，所有的隱藏層節點 (Hidden Node) 皆被視為一致的個體，配合倒傳遞演算法 (Back Propagation Algorithm) 調整權值 (Weight) 以解決問題。然而倒傳遞演算法有訓練效率過低的缺點，並且在整合基因演算法 (Genetic Algorithm, GA) 演化類神經網路結構時，運用此法調整權值，也容易產生結構與權值無法配合的問題。 本論文提出賦予隱藏層節點特定的功能，節點據此調整權值，期望所得的輸出與指定功能的目標值一致。而整體網路的訓練便是以單層的網路訓練為基礎，由底而上漸次完成。並且利用基因演算法，選擇適當的節點功能，安排在適當的位置，讓整個系統得以因應不同的問題，自動產生適合的類神經網路結構。本方法具有隱藏層節點的定位確定，節點功能的再利用，網路訓練透明化與增益演化效率幾項特點。

In traditional Artificial Neural Networks, all hidden nodes are regarded identical, and the nodes adapt weights to solve problems with Back Propagation Algorithm. However, the speed of convergence of Back Propagation is slow, and when we incorporate Genetic Algorithms in evolving Artificial Neural Networks with this method tuning weights, the network architecture and weights will not match well easily. This paper proposed giving hidden nodes specific functions. Then, the nodes change weights to make their outputs similar to the goals of the functions of them. The whole neural network training is based on the single neural networks training and then completes. In addition, we use Genetic Algorithm to select the functions for right nodes. The system will automatically generate a fit Artificial Neural Network for various applications respectively. The primary characteristics of this method include the niche of hidden nodes, the re-usefulness of node functions, the transparency of network training, and the improvement of evolution performance.