% BP 神經網絡用于模式分類
% 使用平臺 - Matlab6.5
% 作者：陸振波，海軍工程大學
% 歡迎同行來信交流與合作，更多文章與程序下載請訪問我的個人主頁
% 電子郵件：luzhenbo@yahoo.com.cn
% 個人主頁：http://luzhenbo.88uu.com.cn

clc
clear
close all

%---------------------------------------------------
% 產生訓練樣本與測試樣本，每一列為一個樣本

P1 = [rand(3,5),rand(3,5)+1,rand(3,5)+2];
T1 = [repmat([1;0;0],1,5),repmat([0;1;0],1,5),repmat([0;0;1],1,5)];    
%                      
% B =repmat Replicate and tile an array SyntaxB = repmat(A,m,n)
% B = repmat(A,[m n])
% B = repmat(A,[m n p...])
% repmat(A,m,n)
% DescriptionB = repmat(A,m,n) creates a large matrix B consisting of an m-by-n tiling of copies of A. The statement repmat(A,n) creates an n-by-n tiling. B = repmat(A,[m n]) accomplishes the same result as repmat(A,m,n). B = repmat(A,[m n p...]) produces a multidimensional (m-by-n-by-p-by-...) array composed of copies of A. A may be multidimensional. repmat(A,m,n) when A is a scalar, produces an m-by-n matrix filled with A's value. This can be much faster than a*ones(m,n) when m or n is large. ExamplesIn this example, repmat replicates 12 copies of the second-order identity matrix, resulting in a "checkerboard" pattern. B = repmat(eye(2),3,4)

% B =
%      1     0     1     0     1     0     1     0
%      0     1     0     1     0     1     0     1
%      1     0     1     0     1     0     1     0
%      0     1     0     1     0     1     0     1
%      1     0     1     0     1     0     1     0
%      0     1     0     1     0     1     0     1 
% The statement N = repmat(NaN,[2 3]) creates a 2-by-3 matrix of NaNs.
%      1     0     1     0     1     0     1     0
%      0     1     0     1     0     1     0     1
%      1     0     1     0     1     0     1     0
%      0     1     0     1     0     1     0     1
%      1     0     1     0     1     0     1     0
%      0     1     0     1     0     1     0     1 
% The statement N = repmat(NaN,[2 3]) creates a 2-by-3 matrix of NaNs.

P2 = [rand(3,5),rand(3,5)+1,rand(3,5)+2];
T2 = [repmat([1;0;0],1,5),repmat([0;1;0],1,5),repmat([0;0;1],1,5)];

%---------------------------------------------------
% 歸一化

[PN1,minp,maxp] = premnmx(P1);
PN2 = tramnmx(P2,minp,maxp);

%---------------------------------------------------
% 設置網絡參數

NodeNum = 10;                   % 隱層節點數 
TypeNum = 3;                    % 輸出維數

TF1 = 'tansig';TF2 = 'purelin'; % 判別函數(缺省值)
%TF1 = 'tansig';TF2 = 'logsig';
%TF1 = 'logsig';TF2 = 'purelin';
%TF1 = 'tansig';TF2 = 'tansig';
%TF1 = 'logsig';TF2 = 'logsig';
%TF1 = 'purelin';TF2 = 'purelin';

net = newff(minmax(PN1),[NodeNum TypeNum],{TF1 TF2});

%---------------------------------------------------
% 指定訓練參數

% net.trainFcn = 'traingd';  % 梯度下降算法
% net.trainFcn = 'traingdm'; % 動量梯度下降算法
% 
% net.trainFcn = 'traingda'; % 變學習率梯度下降算法
% net.trainFcn = 'traingdx'; % 變學習率動量梯度下降算法 
%
% (大型網絡的首選算法 - 模式識別)
% net.trainFcn = 'trainrp';  % RPROP(彈性BP)算法,內存需求最小 
% 
% 共軛梯度算法 
% net.trainFcn = 'traincgf'; % Fletcher-Reeves修正算法
% net.trainFcn = 'traincgp'; % Polak-Ribiere修正算法,內存需求比Fletcher-Reeves修正算法略大
% net.trainFcn = 'traincgb'; % Powell-Beal復位算法,內存需求比Polak-Ribiere修正算法略大
% (大型網絡的首選算法 - 函數擬合,模式識別)
% net.trainFcn = 'trainscg'; % Scaled Conjugate Gradient算法,內存需求與Fletcher-Reeves修正算法相同,計算量比上面三種算法都小很多 
% 
% net.trainFcn = 'trainbfg'; % Quasi-Newton Algorithms - BFGS Algorithm,計算量和內存需求均比共軛梯度算法大,但收斂比較快
% net.trainFcn = 'trainoss'; % One Step Secant Algorithm,計算量和內存需求均比BFGS算法小,比共軛梯度算法略大
% 
% (中小型網絡的首選算法 - 函數擬合,模式識別)
net.trainFcn = 'trainlm';  % Levenberg-Marquardt算法,內存需求最大,收斂速度最快
%
% net.trainFcn = 'trainbr';  % 貝葉斯正則化算法
%
% 有代表性的五種算法為:'traingdx','trainrp','trainscg','trainoss', 'trainlm'

%---------------------%

net.trainParam.show = 1;          % 訓練顯示間隔
net.trainParam.lr = 0.3;            % 學習步長 - traingd,traingdm
net.trainParam.mc = 0.95;           % 動量項系數 - traingdm,traingdx
net.trainParam.mem_reduc = 10;      % 分塊計算Hessian矩陣(僅對Levenberg-Marquardt算法有效)
net.trainParam.epochs = 1000;       % 最大訓練次數
net.trainParam.goal = 1e-8;         % 最小均方誤差
net.trainParam.min_grad = 1e-20;    % 最小梯度
net.trainParam.time = inf;          % 最大訓練時間

%---------------------------------------------------
% 訓練與測試

net = train(net,PN1,T1);             % 訓練

%---------------------------------------------------
% 測試

Y1 = sim(net,PN1);             % 訓練樣本實際輸出
Y2 = sim(net,PN2);             % 測試樣本實際輸出

Y1 = full(compet(Y1));         % 競爭輸出
Y2 = full(compet(Y2));     




% View code for competDefault Topics  compet Competitive transfer function.
%    
%    Syntax
%  
%      A = compet(N)
%      info = compet(code)
%  
%    Description
%    
%      compet is a transfer function.  Transfer functions
%      calculate a layer's output from its net input.
%    
%      compet(N) takes one input argument,
%        N - SxQ matrix of net input (column) vectors.
%      and returns output vectors with 1 where each net input
%      vector has its maximum value, and 0 elsewhere.
%    
%      compet(code) returns information about this function.
%      These codes are defined:
%        'deriv'  - Name of derivative function.
%        'name'   - Full name.
%        'output' - Output range.
%        'active' - Active input range.
%  
%      compet does not have a derivative function.
%    
%    Examples
%  
%      Here we define a net input vector N, calculate the output,
%      and plot both with bar graphs.
%  
%        n = [0; 1; -0.5; 0.5];
%        a = compet(n);
%        subplot(2,1,1), bar(n), ylabel('n')
%        subplot(2,1,2), bar(a), ylabel('a')

%---------------------------------------------------
% 結果統計

Result = ~sum(abs(T1-Y1))                  % 正確分類顯示為1
Percent1 = sum(Result)/length(Result)      % 訓練樣本正確分類率

Result = ~sum(abs(T2-Y2))                  % 正確分類顯示為1
Percent2 = sum(Result)/length(Result)      % 測試樣本正確分類率