用遗传算法优化BP神经网络的Matlab编程实例（转）

frogfish

由于BP网络的权值优化是一个无约束优化问题，而且权值要采用实数编码，所以直接利用Matlab遗传算法工具箱。以下贴出的代码是为一个19输入变量，1个输出变量情况下的非线性回归而设计的，如果要应用于其它情况，只需改动编解码函数即可。

1. 程序一：GA训练BP权值的主函数
   
2. function net=GABPNET(XX,YY)
   
3. %--------------------------------------------------------------------------
   
4. %  GABPNET.m
   
5. %  使用遗传算法对BP网络权值阈值进行优化，再用BP算法训练网络
   
6. %--------------------------------------------------------------------------
   
7. %数据归一化预处理
   
8. nntwarn off
   
9. XX=premnmx(XX);
   
10. YY=premnmx(YY);
    
11. %创建网络
    
12. net=newff(minmax(XX),[19,25,1],{'tansig','tansig','purelin'},'trainlm');
    
13. %下面使用遗传算法对网络进行优化
    
14. P=XX;
    
15. T=YY;
    
16. R=size(P,1);
    
17. S2=size(T,1);
    
18. S1=25;%隐含层节点数
    
19. S=R*S1+S1*S2+S1+S2;%遗传算法编码长度
    
20. aa=ones(S,1)*[-1,1];
    
21. popu=50;%种群规模
    
22. initPpp=initializega(popu,aa,'gabpEval');%初始化种群
    
23. gen=100;%遗传代数
    
24. %下面调用gaot工具箱，其中目标函数定义为gabpEval
    
25. [x,endPop,bPop,trace]=ga(aa,'gabpEval',[],initPpp,[1e-6 1 1],'maxGenTerm',gen,...
    
26.   'normGeomSelect',[0.09],['arithXover'],[2],'nonUnifMutation',[2 gen 3]);
    
27. %绘收敛曲线图
    
28. figure(1)
    
29. plot(trace(:,1),1./trace(:,3),'r-');
    
30. hold on
    
31. plot(trace(:,1),1./trace(:,2),'b-');
    
32. xlabel('Generation');
    
33. ylabel('Sum-Squared Error');
    
34. figure(2)
    
35. plot(trace(:,1),trace(:,3),'r-');
    
36. hold on
    
37. plot(trace(:,1),trace(:,2),'b-');
    
38. xlabel('Generation');
    
39. ylabel('Fittness');
    
40. %下面将初步得到的权值矩阵赋给尚未开始训练的BP网络
    
41. [W1,B1,W2,B2,P,T,A1,A2,SE,val]=gadecod(x);
    
42. net.LW{2,1}=W1;
    
43. net.LW{3,2}=W2;
    
44. net.b{2,1}=B1;
    
45. net.b{3,1}=B2;
    
46. XX=P;
    
47. YY=T;
    
48. %设置训练参数
    
49. net.trainParam.show=1;
    
50. net.trainParam.lr=1;
    
51. net.trainParam.epochs=50;
    
52. net.trainParam.goal=0.001;
    
53. %训练网络
    
54. net=train(net,XX,YY);
    
55. 
    
56. 
    
57. 程序二：适应值函数
    
58. function [sol, val] = gabpEval(sol,options)
    
59. % val - the fittness of this individual
    
60. % sol - the individual, returned to allow for Lamarckian evolution
    
61. % options - [current_generation]
    
62. load data2
    
63. nntwarn off
    
64. XX=premnmx(XX);
    
65. YY=premnmx(YY);
    
66. P=XX;
    
67. T=YY;
    
68. R=size(P,1);
    
69. S2=size(T,1);
    
70. S1=25;%隐含层节点数
    
71. S=R*S1+S1*S2+S1+S2;%遗传算法编码长度
    
72. for i=1:S,
    
73.    x(i)=sol(i);
    
74. end;
    
75. [W1, B1, W2, B2, P, T, A1, A2, SE, val]=gadecod(x);
    
76. 
    
77. 程序三：编解码函数
    
78. function [W1, B1, W2, B2, P, T, A1, A2, SE, val]=gadecod(x)
    
79. load data2
    
80. nntwarn off
    
81. XX=premnmx(XX);
    
82. YY=premnmx(YY);
    
83. P=XX;
    
84. T=YY;
    
85. R=size(P,1);
    
86. S2=size(T,1);
    
87. S1=25;%隐含层节点数
    
88. S=R*S1+S1*S2+S1+S2;%遗传算法编码长度
    
89. % 前R*S1个编码为W1
    
90. for i=1:S1,
    
91.     for k=1:R,
    
92.       W1(i,k)=x(R*(i-1)+k);
    
93.     end
    
94. end
    
95. % 接着的S1*S2个编码（即第R*S1个后的编码）为W2
    
96. for i=1:S2,
    
97.    for k=1:S1,
    
98.       W2(i,k)=x(S1*(i-1)+k+R*S1);
    
99.    end
    
100. end
     
101. % 接着的S1个编码（即第R*S1+S1*S2个后的编码）为B1
     
102. for i=1:S1,
     
103.    B1(i,1)=x((R*S1+S1*S2)+i);
     
104. end
     
105. % 接着的S2个编码（即第R*S1+S1*S2+S1个后的编码）为B2
     
106. for i=1:S2,
     
107.    B2(i,1)=x((R*S1+S1*S2+S1)+i);
     
108. end
     
109. % 计算S1与S2层的输出
     
110. A1=tansig(W1*P,B1);
     
111. A2=purelin(W2*A1,B2);
     
112. % 计算误差平方和
     
113. SE=sumsqr(T-A2);
     
114. val=1/SE; % 遗传算法的适应值

复制代码

上述程序需要调用gaot工具箱

来自：2nsoft.cn

花如月

楼主辛苦了，多谢！

mcx8305

请问frogfish,您能否上传initializega 函数的代码，我的是gads工具箱，没有这个函数。谢谢啦，在线等哦~

mcx8305

mcx8305

[x,endPop,bPop,trace]=ga(aa,'gabpEval',[],initPpp,[1e-6 1 1],'maxGenTerm',gen,...
  'normGeomSelect',[0.09],['arithXover'],[2],'nonUnifMutation',[2 gen 3]);
这一句等号后面的怎么多项都是什么意思啊？我运行后出现：
??? Error using ==> ga
Too many input arguments.

Error in ==> GABPNET at 24
[x,endPop,bPop,trace]=ga(aa,'gabpEval',[],initPpp,[1e-6 1 1],'maxGenTerm',gen,...
谢谢解答！

mcx8305

[x,endPop,bPop,trace]=ga(aa,'gabpEval',[],initPpp,[1e-6 1 1],'maxGenTerm',gen,...
  'normGeomSelect',[0.09],['arithXover'],[2],'nonUnifMutation',[2 gen 3]);
这一句等号后面的怎么多项都是什么意思啊？我运行后出现：
??? Error using ==> ga
Too many input arguments.

Error in ==> GABPNET at 24
[x,endPop,bPop,trace]=ga(aa,'gabpEval',[],initPpp,[1e-6 1 1],'maxGenTerm',gen,...
谢谢解答！

freeman0502609

我用你那程序算出来的为什么没办法收敛呀

ziyanghh152

to mcx8305::


[x,endPop,bPop,trace]=ga(aa,'gabpEval',[],initPpp,[1e-6 1 1],'maxGenTerm',gen,...
  'normGeomSelect',[0.09],['arithXover'],[2],'nonUnifMutation',[2 gen 3]);

去掉那三个省略号！！:loveliness:

tianyoume

每次结果都不完全一致,但误差都应该在合理的范围内.

liujiadong0220

下载下来，好好的研究，可能对我的硕士论文很大的帮助！谢谢lz

chg555

下载下来，好好的研究，感谢感谢....

xxzhao

一点也不懂。新手该如何开始呢？在线等待指教:handshake

frogfish

原帖由 mcx8305 于 2007-8-31 08:35 发表

                               
登录/注册后可看大图

请问frogfish,您能否上传initializega 函数的代码，我的是gads工具箱，没有这个函数。谢谢啦，在线等哦~

1. function [pop] = initializega(num, bounds, evalFN,evalOps,options)
   
2. % function [pop]=initializega(populationSize, variableBounds,evalFN,
   
3. %                           evalOps,options)
   
4. %    initializega creates a matrix of random numbers with
   
5. %    a number of rows equal to the populationSize and a number
   
6. %    columns equal to the number of rows in bounds plus 1 for
   
7. %    the f(x) value which is found by applying the evalFN.
   
8. %    This is used by the ga to create the population if it
   
9. %    is not supplied.
   
10. %
    
11. % pop            - the initial, evaluated, random population
    
12. % populatoinSize - the size of the population, i.e. the number to create
    
13. % variableBounds - a matrix which contains the bounds of each variable, i.e.
    
14. %                  [var1_high var1_low; var2_high var2_low; ....]
    
15. % evalFN         - the evaluation fn, usually the name of the .m file for
    
16. %                  evaluation
    
17. % evalOps        - any options to be passed to the eval function defaults []
    
18. % options        - options to the initialize function, ie.
    
19. %                  [type prec] where eps is the epsilon value
    
20. %                  and the second option is 1 for float and 0 for binary,
    
21. %                  prec is the precision of the variables defaults [1e-6 1]
    
22. 
    
23. % Binary and Real-Valued Simulation Evolution for Matlab GAOT V2
    
24. % Copyright (C) 1998 C.R. Houck, J.A. Joines, M.G. Kay
    
25. %
    
26. % C.R. Houck, J.Joines, and M.Kay. A genetic algorithm for function
    
27. % optimization: A Matlab implementation. ACM Transactions on Mathmatical
    
28. % Software, Submitted 1996.
    
29. %
    
30. % This program is free software; you can redistribute it and/or modify
    
31. % it under the terms of the GNU General Public License as published by
    
32. % the Free Software Foundation; either version 1, or (at your option)
    
33. % any later version.
    
34. %
    
35. % This program is distributed in the hope that it will be useful,
    
36. % but WITHOUT ANY WARRANTY; without even the implied warranty of
    
37. % MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
    
38. % GNU General Public License for more details. A copy of the GNU
    
39. % General Public License can be obtained from the
    
40. % Free Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
    
41. 
    
42. if nargin<5
    
43.   options=[1e-6 1];
    
44. end
    
45. if nargin<4
    
46.   evalOps=[];
    
47. end
    
48. 
    
49. if any(evalFN<48) %Not a .m file
    
50.   if options(2)==1 %Float GA
    
51.     estr=['x=pop(i,1); pop(i,xZomeLength)=', evalFN ';'];  
    
52.   else %Binary GA
    
53.     estr=['x=b2f(pop(i,:),bounds,bits); pop(i,xZomeLength)=', evalFN ';'];
    
54.   end
    
55. else %A .m file
    
56.   if options(2)==1 %Float GA
    
57.     estr=['[ pop(i,:) pop(i,xZomeLength)]=' evalFN '(pop(i,:),[0 evalOps]);'];
    
58.   else %Binary GA
    
59.     estr=['x=b2f(pop(i,:),bounds,bits);[x v]=' evalFN ...
    
60.         '(x,[0 evalOps]); pop(i,:)=[f2b(x,bounds,bits) v];'];  
    
61.     end
    
62. end
    
63. 
    
64. 
    
65. numVars     = size(bounds,1);                 %Number of variables
    
66. rng         = (bounds(:,2)-bounds(:,1))'; %The variable ranges'
    
67. 
    
68. if options(2)==1 %Float GA
    
69.   xZomeLength = numVars+1;                 %Length of string is numVar + fit
    
70.   pop         = zeros(num,xZomeLength);         %Allocate the new population
    
71.   pop(:,1:numVars)=(ones(num,1)*rng).*(rand(num,numVars))+...
    
72.     (ones(num,1)*bounds(:,1)');
    
73. else %Binary GA
    
74.   bits=calcbits(bounds,options(1));
    
75.   xZomeLength = sum(bits)+1;                 %Length of string is numVar + fit
    
76.   pop = round(rand(num,sum(bits)+1));
    
77. end
    
78. 
    
79. for i=1:num
    
80.   eval(estr);
    
81. end

复制代码

frogfish

原帖由 tianyoume 于 2008-3-14 12:01 发表

                               
登录/注册后可看大图

每次结果都不完全一致,但误差都应该在合理的范围内.

初始化每次都不是一样，这是必然的

frogfish

原帖由 xxzhao 于 2008-5-17 18:21 发表

                               
登录/注册后可看大图

一点也不懂。新手该如何开始呢？在线等待指教:handshake

这不动就无能为力了，自己去看看matlab吧