1 简介

针对传统BP神经网络存在学习效率低、收敛速度慢和容易陷入局部极小值的问题,提出一种基于改进的PSO来优化BP神经网络的方法。实验结果表明,该方法较好地解决了传统BP神经网络易陷入局部极小值的问题,提高了算法的收敛速度和稳定性。

2 部分代码

​%%warning('off');% Data Loadingclear;netdata=load('fortest2.mat');netdata=netdata.FinalReady;% Data and Labelnetwork=netdata(:,1:end-1);netlbl=netdata(:,end);% Var Changeinputs = network;targets = netlbl;% Dim SizeInputNum = size(inputs,2);OutputNum = size(targets,2);pr = [-1 1];PR = repmat(pr,InputNum,1);% NN Structure (log-sigmoid transfer function)​NH=5;     % Number of Hidden Layers (more better)​Network1 = newff(PR,[NH OutputNum],{'tansig' 'tansig'});% Train with PSO on Networks WeightsNetwork1 = TrainPSO(Network1,inputs,targets);view(Network1)% Generating Outputs from Our PSO + NN Network Modeloutputs = Network1(inputs');outputs=outputs';% Sizesizenet=size(network);sizenet=sizenet(1,1);% Outputs ErrorMSE=mse(outputs);% Bias Output for Confusion Matrixoutputs=outputs-(MSE*0.1)/2;% Detecting Mislabeled Datafor i=1 : 50            if outputs(i) <= 0.9               out(i)=0;        elseif outputs(i) >= 0.9               out(i)=1;            end;end;for i=51 : 100            if outputs(i) <= 0.9               out(i)=0;        elseif outputs(i) >= 0.9               out(i)=2;            end;end;for i=101 : 150            if outputs(i) <= 0.9               out(i)=0;        elseif outputs(i) >= 0.9               out(i)=3;            end;end;for i=151 : 200            if outputs(i) <= 0.9               out(i)=0;        elseif outputs(i) >= 0.9               out(i)=4;            end;end;for i=201 : 250            if outputs(i) <= 0.9               out(i)=0;        elseif outputs(i) >= 0.9               out(i)=5;            end;end;for i=251 : 300            if outputs(i) <= 0.9               out(i)=0;        elseif outputs(i) >= 0.9               out(i)=6;            end;end;       out1=single(out');% PSO Final Accuracy       psomse=mse(out1,targets);       MSEError=abs(mse(targets)-mse(out1));       cnt=0;       for i=1:sizenet           if out1(i)~= targets(i)               cnt=cnt+1;           end;       end;      fin=cnt*100/ sizenet;      psoacc=(100-fin)-psomse;%%% KNN for Comparison lblknn=netdata(:,end);dataknn=netdata(:,1:end-1);Mdl = fitcknn(dataknn,lblknn,'NumNeighbors',8,'Standardize',1);rng(1); % For reproducibilityknndat = crossval(Mdl);classError = kfoldLoss(knndat);% Predict the labels of the training data.predictedknn = resubPredict(Mdl);ctknn=0;for i = 1 : sizenet(1,1)if lblknn(i) ~= predictedknn(i)    ctknn=ctknn+1;end;end;finknn=ctknn*100/ sizenet;KNN=(100-finknn)-classError;%%% SVM for Comparisontsvm = templateSVM('KernelFunction','polynomial');svmclass = fitcecoc(dataknn,lblknn,'Learners',tsvm);svmerror = resubLoss(svmclass);CVMdl = crossval(svmclass);genError = kfoldLoss(CVMdl);% Predict the labels of the training data.predictedsvm = resubPredict(svmclass);ct=0;for i = 1 : sizenet(1,1)if lblknn(i) ~= predictedsvm(i)    ct=ct+1;end;end;% Compute Accuracyfinsvm=ct*100/ sizenet;SVMAccuracy=(100-finsvm);%% Tree for ComparisonMdl2 = fitctree(dataknn,lblknn);rng(1); % For reproducibilitytreedat = crossval(Mdl2);classErrortree = kfoldLoss(treedat);% Predict the labels of the training data.predictedtree = resubPredict(Mdl2);cttree=0;for i = 1 : sizenet(1,1)if lblknn(i) ~= predictedtree(i)    cttree=cttree+1;end;end;fintree=cttree*100/ sizenet;TREE=(100-fintree)-classErrortree;%% Plots and Results% Confusion Matrixfigure% set(gcf, 'Position',  [50, 100, 1300, 300])subplot(2,2,1)cmsvm = confusionchart(lblknn,predictedsvm);cmsvm.Title = (['SVM Classification =  ' num2str(SVMAccuracy) '%']);subplot(2,2,2)cmknn = confusionchart(lblknn,predictedknn);cmknn.Title = (['KNN Classification =  ' num2str(KNN) '%']);subplot(2,2,3)cmtree = confusionchart(lblknn,predictedtree);cmtree.Title = (['Tree Classification =  ' num2str(TREE) '%']);subplot(2,2,4)cmpso = confusionchart(out1,targets);cmpso.Title = (['PSO-NN Classification =  ' num2str(psoacc) '%']);% Regressionfigureset(gcf, 'Position',  [50, 150, 450, 350])[population2,gof] = fit(targets,out1,'poly4');plot(targets,out1,'o',...    'LineWidth',3,...    'MarkerSize',5,...    'Color',[0.3,0.9,0.2]);    title(['PSO - R =  ' num2str(1-gof.rmse)]);    xlabel('Train Target');    ylabel('Train Output');   hold onplot(population2,'b-','predobs');    xlabel(' Target');    ylabel(' Output');   hold off% ACC and Metrics Resultsfprintf('The SVM Accuracy is = %0.4f.\n',SVMAccuracy)fprintf('The KNN Accuracy is = %0.4f.\n',KNN)fprintf('The Tree Accuracy is = %0.4f.\n',TREE)fprintf('The PSO Accuracy is = %0.4f.\n',psoacc)fprintf('PSO MSE is = %0.4f.\n',MSEError)fprintf('PSO RMSE is = %0.4f.\n',sqrt(MSEError))fprintf('PSO MAE is = %0.4f.\n',mae(targets,out1))​​​​

3 仿真结果

4 参考文献

[1]杨宝华, 叶生波, 戴前颖,等. 一种基于粒子群算法优化BP神经网络的茶叶存储时间分类方法:, 2019.

博主简介：擅长智能优化算法、神经网络预测、信号处理、元胞自动机、图像处理、路径规划、无人机等多种领域的Matlab仿真，相关matlab代码问题可私信交流。

部分理论引用网络文献，若有侵权联系博主删除。

​