matlab-基于MATLAB的ESN网络训练测试仿真-源码

clc; clear; close all; warning off; addpath(genpath(pwd)); %Echo State Network (ESN) %% Parameters actFunc = @ActFuncs.Tanh; alpha=1; % leaking rate betta=1e-8; % regularization coefficient regInput=0.01; % regularization for W_in regHidden=0.98; % regularization for K Sparse=0.3; % percentage of zero weights of K % Network structure inputSize = 1; hiddenSize = 500; outputSize = 1000; samplesSize=10000; trialsSize=samplesSize-outputSize; %% Build the ESN %input layer X= rand(1,samplesSize); W_in= rand(hiddenSize,inputSize)*regInput; %reservoir layer K= randn(hiddenSize) .* randi([0,1],hiddenSize); rhow= max(real(eig(K))); % biggest eigenvalue K= (K./rhow)* regHidden; % update K r= zeros(hiddenSize,1); % reservoir state % Matrice for each time step, saves the reservoier state (with bias & input neurons) r_mat=zeros(trialsSize,hiddenSize+2); % output layer Y0=zeros(outputSize,trialsSize); % teachers for predictions rSquare_train=zeros(outputSize,1); rSquare_test=zeros(outputSize,1); %% Learning Process of the ESN % feed forward for i = 1:length(X) r = (1-alpha)*r + alpha*Tanh(W_in*X(i)+K*r); % Dynamic equation of recurrent network if i > outputSize currInd=i-outputSize; Y0(:,currInd)=X(currInd:i-1)'; % update Y0 r_mat(currInd,:) = [r;X(i);1]; % update reservoir state + input and bias end end % learning of W_out using linear regression C = (1/trialsSize) * (r_mat)' * (r_mat); u = (1/trialsSize)*Y0*r_mat; % ridge regression W_out = (C + betta * eye(size(C, 1))) \ u'; Y = (r_mat*W_out)'; % output of reservoir (Prediction) %% Computational part % R square Score per neuron for neuron = 1 : outputSize rSquare_train(neuron) = corr(Y(neuron, :)', Y0(neuron, :)') .^ 2; end memoryCapacity=sum(rSquare_train); disp(['Training Memory Capacity is : ' num2str(memoryCapacity)]) %% ESN Testing X_test = randn(1,samplesSize); Y0_test = zeros(outputSize,trialsSize); r_mat_test = zeros(trialsSize,hiddenSize+2); % feed forword for i = 1:length(X) r = (1-alpha)*r + alpha*Tanh(W_in*X(i)+K*r); if i > outputSize currInd=i-outputSize; Y0_test(:,currInd)=X(currInd:i-1)'; % update Y0 r_mat_test(currInd,:) = [r;X(i);1]; % update reservoir state + input and bias end end Y_test = (r_mat_test*W_out)'; % R square Score per neuron for neuron = 1 : outputSize rSquare_test(neuron) = corr(Y_test(neuron, :)', Y0_test(neuron, :)') .^ 2; end memoryCapacity_test=sum(rSquare_test); disp(['Test Memory Capacity is : ' num2str(memoryCapacity_test)]) %% Plot Result plotResult(outputSize, rSquare_train,rSquare_test);