韶关学院学报
韶關學院學報
소관학원학보
Journal of Shaoguan University(Social Science Edition)
2013年
12期
18~21
,共null页
BP神经网络 权值 阈值 初始化
BP神經網絡 權值 閾值 初始化
BP신경망락 권치 역치 초시화
BP neural network ; weights ; threshold ; initialization
提出一种等比例系数型BP神经网络权值和阈值的初始化方法,该方法可将S型传递函数的自变量转化到区间[0,1）上,从而提高网络的收敛性能.从理论上证明了该方法的可行,并通过非线性函数y=x12＋x22的拟合实验,证明了该方法是有效的.
提齣一種等比例繫數型BP神經網絡權值和閾值的初始化方法,該方法可將S型傳遞函數的自變量轉化到區間[0,1）上,從而提高網絡的收斂性能.從理論上證明瞭該方法的可行,併通過非線性函數y=x12＋x22的擬閤實驗,證明瞭該方法是有效的.
제출일충등비례계수형BP신경망락권치화역치적초시화방법,해방법가장S형전체함수적자변량전화도구간[0,1）상,종이제고망락적수렴성능.종이론상증명료해방법적가행,병통과비선성함수y=x12＋x22적의합실험,증명료해방법시유효적.
The paper presents an initialized method of equal-scaling coefficient type based on BP neural network weights and thresholds, which can convert the independent variable of sigmoid transfer function to the interval I0,1 ）, thereby improving the convergence performance of the network. It is theoretically proved that the method is feasible, and it shows that the method is effective proved by fitting experiments on the non-linear functiony=xl2＋X22.
