基于三次收敛LM算法的神经网络研究

ON NEURAL NETWORK BASED ON THREE CONVERGENCES LM ALGORITHM

为了解决传统神经网络实际应用中计算复杂、耗时过长等问题,在LM(Levenberg-Marquardt)算法的基础上,结合数学最优化理论,找出三次收敛的改进型LM算法,且将其应用于BP神经网络。利用一组火灾现场数据,通过Matlab仿真对改进型LMBP算法和标准LMBP算法从收敛时间和仿真曲线拟合度两方面进行比较。结果表明改进型LMBP算法在收敛时间和拟合度两方面都有更好的效果,且该算法具有一般性,可以通过获取国民生产中各种应用场景的样本,采用该算法进行预测,更好地指导生产。

In order to solve the practical problems of traditional neural network application in computation complexity and long time consu- ming, on the basis of LM algorithm and combined with mathematical optimisation theory, we find out the improved LM algorithm with three convergences and apply it to BP neural network. Moreover, by using a set of data got from the scene of the fire, we compare the improved LMBP algorithm and the standard LMBP algorithm through Matlab simulation in two aspects of convergence time and simulation curve fitting. Experiments show that the improved LMBP algorithm has better effect than the standard LMBP algorithm in these two aspects. The improved LMBP algorithm also has the generality, it can be used to predict for better guiding the production by obtaining the samples of various scenari- os in national product.