💥1 概述
该文提出了一种新的长度可变极限学习机。该方法旨在提高单隐藏层前馈神经网络（SLFN）在丰富的动态不平衡数据下的学习性能。粒子群优化 （PSO） 涉及增量学习期间的超参数调整和更新。该算法使用燃气涡轮发动机C-MAPSS（商用模块化航空推进系统仿真）数据集的子集进行评估，并与其衍生物进行比较。结果表明，新算法具有较好的学习效果。
文献来源：
📚2 运行结果
部分代码：
function [net] = OP_W_LCI_ELM(xtr,ytr,xts,yts,sample,Options)
% Length changeabale incremental ELM :
% hyper parameters
k=Options.k;
lambda=Options.lambda;
MaxHiddenNeurons=Options.MaxHiddenNeurons;
epsilon=Options.epsilon;
ActivationFunType=Options.ActivationFunType;
C=Options.C;
Weighted=Options.Weighted;
maxite=Options.maxite;
epsilonPSO=Options.epsilonPSO;
%%%%%%%%%%%%%%%%
[~,InputDimension] = size(xtr);% get dimensions
start_time_training = cputime;% calculating time
W = [];% initialize input weights
b = [];% initialize hidden layer biases
TrainingAccuracy = [];% initialize
OutputWeights = [];% initialize
num = 0; % initialize number of hidden nodes
count = 0;% initialize ITERATION COUNTER
E=[]; % nitialize vector of RMSEs by iteration
Ets=[]; % initialize vector of RMSEs of testing
Tolerance=1;% initialize tolerance value
%Training phase
node=0;
while num < MaxHiddenNeurons & Tolerance > epsilon
len = floor(k*exp(-lambda*count))+1; %calculate the number of newly added hidden nodes for this loop
wi = rand(InputDimension,len)*2-1;
W = [W wi];
switch ActivationFunType
case 'sig'
bi = rand(1,len)*2-1;
b = [b bi];
H = activeadd(xtr,W,b);
case 'radbas'
bi = rand(1,len)*0.5;
b = [b bi];
H = activerbf(xtr,W,b);
otherwise
end
% particl swarm optimization for best regularization and weightes
% Adjustments
% dfine initial population
population=[C Weighted];
% random search
[~,population,~,fit_behavior]=PSO(population,H,ytr,Options);
% save best solutions
Reg(count+1,:)=population;
% determine beta
[beta] = R(H,ytr, population);
% calculate target
f = H*beta;
%update the output function
OutputWeights = beta; %update the output weights
TrainingAccuracy = sqrt(mse(ytr-f));
E=[E TrainingAccuracy];% save training accuraccy
if count>1
Tolerance= sqrt(mse(E(count)-E(count-1))); % tolerance function
Tol(count-1,1)=Tolerance;
end
count = count+1;% counter
% testing phase
[O,TestingAccuracy,TestingTime]=ELMtest(xts,yts,ActivationFunType,W,b,OutputWeights);
% testing accuracy
Ets=[Ets TestingAccuracy];
% number of used hiddden nodes
num = num+len;
% used hidden nodes in each step
nodes(count)=node+len;
% stor
node=nodes(count);
end
end_time_training = cputime;
TrainingTime = end_time_training - start_time_training;
clear wi bi H beta ;
% Testing phase
% sample
switch ActivationFunType
case 'sig'
y_sample = activeadd(sample,W,b)*OutputWeights;
case 'radbas'
y_sample = activerbf(sample,W,b)*OutputWeights;
otherwise
end
y_sample=scaledata(y_sample,7,125);
num=num-1;
% performances
[SCORE,S,d,er]=Score(O,yts);
% stor results
%%%
net.E=E;
net.Ets=Ets;
net.num=num;
net.nodes=nodes;
net.Tol=Tol;
net.Tr_time=TrainingTime;
net.Ts_time=TestingTime;
net.Tr_acc=TrainingAccuracy;
net.Ts_acc=TestingAccuracy;
net.Opt_solution=population;
net.Score=SCORE;
net.S=S;
net.d=d;
net.er=er;
net.reg=Reg;
net.y_sample=y_sample;
net.fit_behavior=fit_behavior;
end
🎉3 参考文献
[1] Y. X. Wu, D. Liu, and H. Jiang, “Length-Changeable Incremental Extreme Learning Machine,” J. Comput. Sci. Technol., vol. 32, no. 3, pp. 630–643, 2017.
[2] A. Saxena, M. Ieee, K. Goebel, D. Simon, and N. Eklund, “Damage Propagation Modeling for Aircraft Engine Prognostics,” Response, 2008.
[3] M. N. Alam, “Codes in MATLAB for Particle Swarm Optimization Codes in MATLAB for Particle Swarm Optimization,” no. March, 2016.
[4] J. Cao, K. Zhang, M. Luo, C. Yin, and X. Lai, “Extreme learning machine and adaptive sparse representation for image classification,” Neural Networks, vol. 81, no. 61773019, pp. 91–102, 2016.