基于NSGA-Ⅱ算法的BP神经网络优化研究

Modeling of penicillin ferment process based on the combination of non-dominated sorting genetic algorithm Ⅱ and error back propagation network

针对BP(Back Propagation)网络训练时,会遇到陷入局部极小点、收敛速度慢等问题,提出将BP训练看成多目标寻优过程,以网络输出节点的误差最小作为并行搜索的多个目标,通过带精英策略的非支配排序遗传算法(NSGA-Ⅱ,Non-Dominated Sorting Genetic Algorithm Ⅱ)对BP网络的初始权值和阈值进行优化,给出了应用该方法的步骤。通过仿真验证,这种NSGA-Ⅱ&BP算法对一个单输入双输出非线性系统进行逼近,能克服BP网络训练的缺陷,且所建模型对检测样本的拟合程度比单独BP网络的效果要好。

To overcome the normal problems of BP, such as it is easy to be trapped in local minima and its convergence speed is slow when error back propagation network training. A method is put forward to bring the concept of multi-objective optimization in back propagation network, and regard the output errors of BP network as the multi-objectives to minimize in parallel. Using nondominated sorting genetic algorithm Ⅱ to optimize the initial weights and threshold values of BP. The procedure of this method is given. The simulation result indicated that the combination method would get a better approaching effect Jn modeling of a single input-double outputs system.