%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% Matlab script to perform cross validation on numeric data from agricult.
% purpose: data mining with ANNs/MLPs 
% see http://blog.georgruss.de/?p=62
% -----
% script to generate plots and determine optimal network
% for agriculture data, accompanies a paper submitted for review
% at ICDM2008 http://www.data-mining-forum.de/icdm2008.php
% -----
% Georg Ru\ss
% russ@iws.cs.uni-magdeburg.de
% 2008-01-18
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

%% Preparation steps, workspace

% clean workspace
clear all;

% set clock
clock_start = clock;
% seed random for reproducible results
rand('seed',1);

%pause;

% change paths to wherever your data is located and readable to matlab
% uses script readColData from
% http://web.cecs.pdx.edu/~gerry/MATLAB/plotting/loadingPlotData.html#colHeadings
[label,id_column,data]=readColData('sorted_all',10,10,1);

%% data specific stuff
% generate three data sets for the nnet to play with
% one: N1, Yield2003, EM38 -- target: Yield2004
% two: N1, Yield2003, EM38, N2, REIP32 -- target: Yield2004
% three: N1, Yield2003, EM38, N2, REIP32, N3, REIP49 -- target: Yield2004
% works by eliminating the respective columns from the 'data' matrix above

set_1 = data;
set_1(:,7) = [];
set_1(:,5) = [];
set_1(:,4) = [];
set_1(:,3) = [];
set_1(:,2) = [];

set_2 = data;
set_2(:,7) = [];
set_2(:,5) = [];
set_2(:,3) = [];

set_3 = data;
set_3(:,7) = [];

%% store dimensions of data for possible later loops
%[number_of_examples_cv,number_of_attributes_kv] = size(crossv);

%% Prepare the data for the neural net
% put input attributes and target values into suitable matrices
% requires selecting and transposing the raw input data

size_set_1 = size(set_1);
size_set_2 = size(set_2);
size_set_3 = size(set_3);

set_1_examples = transpose(set_1(:,1:size_set_1(1,2)-1));
set_1_targets = transpose(set_1(:,size_set_1(1,2)));
set_2_examples = transpose(set_2(:,1:size_set_2(1,2)-1));
set_2_targets = transpose(set_2(:,size_set_2(1,2)));
set_3_examples = transpose(set_3(:,1:size_set_3(1,2)-1));
set_3_targets = transpose(set_3(:,size_set_3(1,2)));

% max size of first hidden layer
max_size_first_layer = 32;
% max size of second hidden layer
max_size_second_layer = 32;

% number of iterations
number_of_iterations = 250;

nets_perf_collection = zeros(number_of_iterations,3);
%net_1_perf_collection = zeros(max_size_first_layer,max_size_second_layer);
%net_2_perf_collection = zeros(max_size_first_layer,max_size_second_layer);
%net_3_perf_collection = zeros(max_size_first_layer,max_size_second_layer);

%create network
for i = 1:number_of_iterations;
    %for j = 2:max_size_second_layer;
        
        % %%%%%%%%%%%%%
        % network for first data set 
        net_1 = newff(set_1_examples,set_1_targets,[16 16]);
        net_1.trainparam.min_grad = 0.001;
        net_1.trainParam.epochs = 100000;
        net_1.trainParam.lr = 0.5;
        net_1.trainParam.show = NaN;
        net_1.divideFcn = 'dividerand';
        % training step
        [net_1_trained,net_1_tr] = train(net_1,set_1_examples,set_1_targets);
        % determine mse on test data (from network training)
        net_1_perf = net_1_tr.tperf(1,size(net_1_tr.tperf,2));
        nets_perf_collection(i,1) = sqrt(net_1_perf);
       
        % %%%%%%%%%%%%%
        % network for second data set
        net_2 = newff(set_2_examples,set_2_targets,[16 16]);
        net_2.trainparam.min_grad = 0.001;
        net_2.trainParam.epochs = 100000;
        net_2.trainParam.lr = 0.5;
        net_2.trainParam.show = NaN;
        net_2.divideFcn = 'dividerand';
        % training step
        [net_2_trained,net_2_tr] = train(net_2,set_2_examples,set_2_targets);
        % determine mse on test data (from network training)
        net_2_perf = net_2_tr.tperf(1,size(net_2_tr.tperf,2));
        nets_perf_collection(i,2) = sqrt(net_2_perf);

        % %%%%%%%%%%%%%
        % network for third data set
        net_3 = newff(set_3_examples,set_3_targets,[16 16]);
        net_3.trainparam.min_grad = 0.001;
        net_3.trainParam.epochs = 100000;
        net_3.trainParam.lr = 0.5;
        net_3.trainParam.show = NaN;
        net_3.divideFcn = 'dividerand';
        % network training
        [net_3_trained,net_3_tr] = train(net_3,set_3_examples,set_3_targets);
        % determine mse on test data (from network training)
        net_3_perf = net_3_tr.tperf(1,size(net_3_tr.tperf,2));
        nets_perf_collection(i,3) = sqrt(net_3_perf);
        
        
        i
     %   j
   % end
end

%%
clf;

nets_perf_collection_1 = nets_perf_collection(:,1);
nets_perf_collection_2 = nets_perf_collection(:,2);
nets_perf_collection_3 = nets_perf_collection(:,3);

plot(nets_perf_collection_1,'-or');
hold on;
plot(nets_perf_collection_2,'-xb');
plot(nets_perf_collection_3,'-+g');
title('Absolute Error on Different Data Sets');
h = legend('Error FT1','Error FT2','Error FT3',1);
set(h,'Interpreter','none')

% generate difference plots manually from the following data
%diff_net_1_net_2 = net_1_perf_collection - net_2_perf_collection;
%diff_net_2_net_3 = net_2_perf_collection - net_3_perf_collection;
%diff_net_1_net_3 = net_1_perf_collection - net_3_perf_collection;
 
duration = etime(clock, clock_start)