CAJ | 학술논문

针对LM （Levenberg-Marquardt）算法的缺陷，提出一种使用改进的遗传算法和LM算法优化神经网络的混合学习算法（GA-LMbp）。该算法先通过改进的遗传算法粗调得到一组全局最优近似解（即BP网络的初始权值和阈值），再以该近似解为初值，用LM算法优化BP网络进行图像压缩处理。实验结果表明，新算法提高了网络的学习能力和收敛速度，避免了LMbp陷入平坦区或局部极小值。
침대LM （Levenberg-Marquardt）산법적결함，제출일충사용개진적유전산법화LM산법우화신경망락적혼합학습산법（GA-LMbp）。해산법선통과개진적유전산법조조득도일조전국최우근사해（즉BP망락적초시권치화역치），재이해근사해위초치，용LM산법우화BP망락진행도상압축처리。실험결과표명，신산법제고료망락적학습능력화수렴속도，피면료LMbp함입평탄구혹국부겁소치。
To compensate the defects of LM (Levenberg-Marquardt) algorithm,an hybrid learning algorithm (GA-LMbp) in which improved genetic algorithm and LM algorithm were proposed and were used to optimizing the neural network. Firstly,the algorithm retained a set of global optimal approximate solution (initial weights and threshold values of BP network) through being improved genetic algorithm. Then,BP network were optimized by LM algorithm with the ap-proximate solution as initial values, and the BP network were used for image compression. The experimental results showed that the new algorithm improves the learning ability and convergence speed of the network, and it succeeds in avoiding LMbp sinking into flat area or local minimum value.