% Input selection based on KNNC and LOO

clear all;
vsdOpt=vsdOptSet;

fprintf('Loading DS.mat...\n');
load DS.mat				% Load the dataset
% ====== Down sample the training data
fprintf('Down sample DS with a rate of %d...\n', vsdOpt.downSampleRate);
dataSize=size(DS.input, 2);
ratio=round(dataSize/vsdOpt.maxDataSize4inputSelect);
DS.input=DS.input(:, 1:ratio:end);
DS.output=DS.output(:, 1:ratio:end);
fprintf('Actual data size for input selection = %d\n', size(DS.input,2));
%DS.input=inputNormalize(DS.input);	% Input (feature) normalization

plotOpt=1;				% Plot the result of input selection
classifier='knncLoo';			% Use knncLoo.m for input selection
knncPrm.k=1;				% Parameters for the classifier
fprintf('inputName=%s\n', cell2str(DS.inputName));	% Print the input names
switch vsdOpt.inputSelectionMethod
	case 'sequential'
		% ====== Sequential forward selection
		[bestInputIndex, bestRecogRate, allSelectedInput, allRecogRate, elapsedTime]=inputSelectSequential(DS, vsdOpt.selectedInputNum, classifier, knncPrm, plotOpt);
	case 'exhaustive'
		% ====== Exhaustive search
		[bestInputIndex, bestRecogRate, allSelectedInput, allRecogRate, elapsedTime]=inputSelectExhaustive(DS, vsdOpt.selectedInputNum, classifier, knncPrm, plotOpt);
	otherwise
		disp('Unknown method!');
end

fprintf('Recognition rate = %f\n', bestRecogRate*100);
fprintf('bestInputIndex = %s\n', mat2str(bestInputIndex));
fprintf('bestInputIndex = %s\n', cell2str(DS.inputName(bestInputIndex)));

[allRecogRate, index]=sort(allRecogRate);
allSelectedInput=allSelectedInput(index);
figure; inputSelectPlot(allRecogRate*100, allSelectedInput, DS.inputName, mfilename);
fprintf('Save the indices of the selected features to bestInputIndex.mat...\n');
save bestInputIndex bestInputIndex