文化進化演算法於類神經網路的訓練

Cultural Evolution Algorithm for Artificial Neural Network Learning

多元文化種的演變過程包括初始形成、各文化種間的特性傳遞、感染以及學習的行為等階段。本文仿多元文化種的演化概念為架構，提出族群式演算法，稱為文化進化演算法。利用此演算法訓練類神網路，首先應用於非線性函數之預測模式建構，并與倒傳遞演算法比較，所得訓練與測試誤差值皆小於倒傳遞演算法。最後，利用此演算法於類神網路之船舶穩度多數預測模型建構訓練，且與回歸分析法比較，結果顯示預測的誤差皆小於回歸分析法。 The process of cultural evolution includes the initial good habitat, the transmission, influence and learning behaviors between cultural species. Borrowing from the concept of multi-cultural population evolution, this paper proposes a hybrid population-based evolution algorithm, named Cultural Evolution Algorithm, to find the global optimum. When applied to neural network training for modeling a prediction model of a nonlinear function, this algorithm achieves trained errors less than the backward propagation algorithm. Finally, this algorithm is used to train a neural network to establish the prediction model of the vessel stability parameters. Compared with the regression analysis, our algorithm gives smaller predicted errors.