BP神经网络逼近非线性函数.docx

1、BP神经网络逼近非线性函数应用BP神经网络逼近非线性函一、实验要求1、逼近的非线性函数选取为y=sin(x1)+cos(x2)，其中有两个自变量即x1,x2,一个因变量即y。2、逼近误差learnGoal) collectHiddenOut = logsig(HH*input_train); %计算隐含层输出 hiddenOut = collectHiddenOut ones(361,1); networkOut = OO*hiddenOut; %计算网络输出 error = output_train-networkOut; %计算误差 %利用目的函数，判断是否完毕循环 aimJ = sums

2、qr(error) if (aimJlearnGoal) break; end %统计训练次数 trainNum = trainNum+1; %权值阈值调整因子 factor2 = error; factor1 = w2*factor2.*collectHiddenOut.*(1-collectHiddenOut); %调整权值和阈值调节量 dHH = factor1*input_train; dOO = factor2*hiddenOut; %权值阈值调整 if (trainNum3) HH = HH + learnSpeed*dHH; OO = OO + learnSpeed*dOO; c

3、ollectHH = collectHH HH; collectOO = collectOO OO; w1 = HH(:,1:inputnum); b1 = HH(:,1+inputnum); w2 = OO(:,1:hiddennum); b2 = OO(:,1+hiddennum); else %附加动量法 HH = HH + learnSpeed*dHH + 0.94*(collectHH(:,(trainNum-2)*3+1):(trainNum-2)*3+3)-collectHH(:,(trainNum-3)*3+1):(trainNum-3)*3+3); OO = OO + lea

4、rnSpeed*dOO + 0.94*(collectOO(1,(trainNum-2)*10+1):(trainNum-2)*10+10)-collectOO(1,(trainNum-3)*10+1):(trainNum-3)*10+10); collectHH = collectHH HH; collectOO = collectOO OO; w1 = HH(:,1:inputnum); b1 = HH(:,1+inputnum); w2 = OO(:,1:hiddennum); b2 = OO(:,1+hiddennum); end %训练数据测试，计算最大误差率 hiddenOut_t

5、est = logsig(HH*input_train); %参数修改后的隐含层输出 network_test = w2*hiddenOut_test+repmat(b2,1,361); %预测结果 rate = (output_train-network_test)./output_train; %误差率 max_rate = max(abs(rate); %误差率最大值end%显示测试结果%标准函数图像y=sin(x1)+cos(x2)x,y = meshgrid(-4.5:0.1:4.5,-4.5:0.1:4.5);z = sin(x)+cos(y);figure(1)mesh(x,y,

6、z)xlabel(x1);ylabel(x2);zlabel(y);%网络图t1=linspace(min(input(:,1),max(input(:,1);t2=linspace(min(input(:,2),max(input(:,2);X,Y=meshgrid(t1,t2);Z=griddata(input(:,1),input(:,2),network_test,X,Y);figure(2)mesh(X,Y,Z)xlabel(Input1);ylabel(Input2);zlabel(Output);%绘制误差曲线t3=linspace(min(input(:,1),max(input(:,1);t4=linspace(min(input(:,2),max(input(:,2);X1,X2=meshgrid(t3,t4);E=griddata(input(:,1),input(:,2),error,X1,X2);figure(3)mesh(X1,X2,E)xlabel(Input1);ylabel(Input2);zlabel(error);