OnlineSVR Matlab 2006b Code.zip_SVR_online svr_svr matlab_增量

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% % ONLINE SUPPORT VECTOR REGRESSION % % Copyright 2006 - Francesco Parrella % % % % This program is distributed under the terms of the GNU License % %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% % Unlearning of a sample function [SVR, Flops] = Unlearn (SVR, CIndex) % Inizialization Error = SVR.Epsilon / 10; Flops = 0; % CASE 0: Right classified sample if (any(find(SVR.RemainingSetIndexes==CIndex))) ShowMessage(SVR, '> Case 0: the sample was removed from the remaining set.',2); RIndex = find(SVR.RemainingSetIndexes==CIndex); SVR.RemainingSetIndexes = [SVR.RemainingSetIndexes(1:(RIndex-1),1); SVR.RemainingSetIndexes((RIndex+1):RemainingSetElementsNumber(SVR),1)]; SVR.X = [SVR.X(1:(CIndex-1),:); SVR.X((CIndex+1):SVR.SamplesTrainedNumber,:)]; SVR.Y = [SVR.Y(1:(CIndex-1),1); SVR.Y((CIndex+1):SVR.SamplesTrainedNumber,1)]; SVR.Weights = [SVR.Weights(1:(CIndex-1),1); SVR.Weights((CIndex+1):SVR.SamplesTrainedNumber,1)]; SVR.SupportSetIndexes(SVR.SupportSetIndexes>CIndex) = SVR.SupportSetIndexes(SVR.SupportSetIndexes>CIndex)-1; SVR.ErrorSetIndexes(SVR.ErrorSetIndexes>CIndex) = SVR.ErrorSetIndexes(SVR.ErrorSetIndexes>CIndex)-1; SVR.RemainingSetIndexes(SVR.RemainingSetIndexes>CIndex) = SVR.RemainingSetIndexes(SVR.RemainingSetIndexes>CIndex)-1; SVR.SamplesTrainedNumber = SVR.SamplesTrainedNumber - 1; if (SVR.SamplesTrainedNumber==0) SVR.Bias = 0; end return; % Remove the sample from its set else % Error Set if (any(find(SVR.ErrorSetIndexes==CIndex))) EIndex = find(SVR.ErrorSetIndexes==CIndex); SVR.ErrorSetIndexes = [SVR.ErrorSetIndexes(1:(EIndex-1),1); SVR.ErrorSetIndexes((EIndex+1):ErrorSetElementsNumber(SVR),1)]; % SupportSet else SIndex = find(SVR.SupportSetIndexes==CIndex); SVR.SupportSetIndexes = [SVR.SupportSetIndexes(1:(SIndex-1),1); SVR.SupportSetIndexes((SIndex+1):SupportSetElementsNumber(SVR),1)]; SVR.R = RemoveSampleFromR(SVR, SIndex); end end % Inizializations Flag = -1; nIterations = 0; SampleRemoved = 0; % Find the initial margin H = Margin(SVR, SVR.X, SVR.Y); while (~SampleRemoved) % Iteration Count Flops = Flops + 1; if (Flops > (SVR.SamplesTrainedNumber+1)*100) SVR.Verbosity = 3; ShowDetails(SVR, CIndex, H); save 'Error' 'SVR'; error('Loop Found. '); end % Find Beta and Gamma [Beta, Gamma] = FindBetaAndGamma(SVR, CIndex); % Find the samples variations [Lc, Ls, Le, Lr] = FindUnlearningMinVariation (SVR, H, Beta, Gamma, CIndex); ShowVariations(SVR, CIndex, H, Beta, Gamma, Lc, Lc, Ls, Le, Lr); % Find the minimum variation [MinLs, MinLsIndex] = min (abs(Ls)',[],1); [MinLe, MinLeIndex] = min (abs(Le)',[],1); [MinLr, MinLrIndex] = min (abs(Lr)',[],1); [MinValue, Flag] = min ([abs(Lc); abs(MinLs); abs(MinLe); abs(MinLr)],[],1); if (MinValue==inf) error('No weights to modify!',3); end switch (Flag) case 1 MinValueWithSign = Lc; case 2 MinValueWithSign = Ls(MinLsIndex); case 3 MinValueWithSign = Le(MinLeIndex); case 4 MinValueWithSign = Lr(MinLrIndex); end % Find the new variation DeltaC = MinValueWithSign; % Update Weights and Bias if (SupportSetElementsNumber(SVR)>0) SVR.Weights(CIndex) = SVR.Weights(CIndex) + DeltaC; Delta = Beta*DeltaC; SVR.Bias = SVR.Bias + Delta(1); SVR.Weights(SVR.SupportSetIndexes) = SVR.Weights(SVR.SupportSetIndexes) + Delta(2:length(Delta)); H = H + Gamma*DeltaC; else SVR.Bias = SVR.Bias + DeltaC; H = H + DeltaC; end switch (Flag) case 1 % The sample alpha reaches 0 ShowMessage(SVR, '> Case 1: the sample alpha becomes 0', 2); SVR.X = [SVR.X(1:(CIndex-1),:); SVR.X((CIndex+1):SVR.SamplesTrainedNumber,:)]; SVR.Y = [SVR.Y(1:(CIndex-1),1); SVR.Y((CIndex+1):SVR.SamplesTrainedNumber,1)]; SVR.Weights = [SVR.Weights(1:(CIndex-1),1); SVR.Weights((CIndex+1):SVR.SamplesTrainedNumber,1)]; H = [H(1:(CIndex-1),1); H((CIndex+1):SVR.SamplesTrainedNumber,1)]; SVR.SupportSetIndexes(SVR.SupportSetIndexes>CIndex) = SVR.SupportSetIndexes(SVR.SupportSetIndexes>CIndex)-1; SVR.ErrorSetIndexes(SVR.ErrorSetIndexes>CIndex) = SVR.ErrorSetIndexes(SVR.ErrorSetIndexes>CIndex)-1; SVR.RemainingSetIndexes(SVR.RemainingSetIndexes>CIndex) = SVR.RemainingSetIndexes(SVR.RemainingSetIndexes>CIndex)-1; SVR.SamplesTrainedNumber = SVR.SamplesTrainedNumber - 1; if (SVR.SamplesTrainedNumber==1 & ErrorSetElementsNumber(SVR)>0) SVR.ErrorSetIndexes = []; SVR.RemainingSetIndexes = [1]; SVR.Weights(1) = 0; SVR.Bias = Margin(SVR, SVR.X(1,:),SVR.Y(1)); H(1) = 0; end if (SVR.SamplesTrainedNumber==0) SVR.Bias = 0; end SampleRemoved = 1; case 2 % Move Sample from SupportSet to ErrorSet or RemainingSet WeightsValue = SVR.Weights(SVR.SupportSetIndexes(MinLsIndex)); if (abs(WeightsValue)<abs(SVR.C-abs(WeightsValue))) SVR.Weights(SVR.SupportSetIndexes(MinLsIndex)) = 0; else SVR.Weights(SVR.SupportSetIndexes(MinLsIndex)) = sign(SVR.Weights(SVR.SupportSetIndexes(MinLsIndex)))*SVR.C; end SIndex = MinLsIndex; if (SVR.Weights(SVR.SupportSetIndexes(MinLsIndex))==0) % Move Sample from SupportSet to RemainingSet ShowMessage(SVR, ['> Case 3a: move sample ' num2str(SVR.SupportSetIndexes(MinLsIndex)) ' from support to remaining set' ' (Var=' num2str(DeltaC) ')'],2); SVR.RemainingSetIndexes(RemainingSetElementsNumber(SVR)+1,1) = SVR.SupportSetIndexes(MinLsIndex,1); SVR.SupportSetIndexes = [SVR.SupportSetIndexes(1:(MinLsIndex-1),1); SVR.SupportSetIndexes((MinLsIndex+1):SupportSetElementsNumber(SVR),1)]; SVR.R = RemoveSampleFromR(SVR, SIndex); elseif (abs(SVR.Weights(SVR.SupportSetIndexes(MinLsIndex)))== SVR.C) % Move Sample from SupportSet to ErrorSet ShowMessage(SVR, ['> Case 3b: move sample ' num2str(SVR.SupportSetIndexes(MinLsIndex)) ' from support to error set' ' (Var=' num2str(DeltaC) ')'],2); SVR.ErrorSetIndexes(ErrorSetElementsNumber(SVR)+1,1) = SVR.SupportSetIndexes(MinLsIndex,1); SVR.SupportSetIndexes = [SVR.SupportSetIndexes(1:(MinLsIndex-1),1); SVR.SupportSetIndexes((MinLsIndex+1):SupportSetElementsNumber(SVR),1)]; SVR.R = RemoveSampleFromR(SVR, SIndex); else SVR.Weights(SVR.SupportSetIndexes(MinLsIndex))