提出了一种改进遗传算法(GA)和BP算法结合的神经网络模型优化方案。首先采用自适应交叉概率和变异概率的遗传算法优化BP网络的权值,在进化结束时,能够寻到全局最优点附近的点;在遗传算法搜索结果的基础上,利用局部寻优能力较强的动量BP算法,从此点出发,进行局部搜索,进而达到网络的训练目标。仿真实验结果表明,在大庆市2000年到2004年6月降雨量的预测方面,遗传算法与BP算法结合的模型预测误差平均为39.13%,标准BP算法的模型预测误差平均为194.66%。说明GA—BP算法模型预报精度较高,预测能力得到了改进。
A BP network model based on improved GA is presented. Adopted the adaptive cross & mutation probability, the GA is to optimize the weight of network. At the end of evolution, a solution near the global optimum will be found. Then the global optimum in the local area can be found by adopting momentum BP algorithm. The simulation shows this mixed method(GA-BP model) is more precise (average error 39.13%)than pure BP model(average error 194.66%)in terms of June rainfall prediction during 2000 to 2004 in Da Qing. Consequently, the prediction ability of mixed algorithm is much improved.