遗传算法（GA）优化后RBF神经网络优化分析（Matlab代码实现）

目录

1 遗传算法

2 RBF神经网络

3 Matlab代码实现

4 结果 

---

1 遗传算法

2 RBF神经网络

RBF神将网络是一种三层神经网络，其包括输入层、隐层、输出层。从输入空间到隐层空间的变换是非线性的，而从隐层空间到输出层空间变换是线性的。流图如下：

                          

RBF网络的基本思想是：用RBF作为隐单元的“基”构成隐含层空间，这样就可以将输入矢量直接映射到隐空间，而不需要通过权连接。当RBF的中心点确定以后，这种映射关系也就确定了。而隐含层空间到输出空间的映射是线性的，即网络的输出是隐单元输出的线性加权和，此处的权即为网络可调参数。其中，隐含层的作用是把向量从低维度的p映射到高维度的h，这样低维度线性不可分的情况到高维度就可以变得线性可分了，主要就是核函数的思想。

这样，网络由输入到输出的映射是非线性的，而网络输出对可调参数而言却又是线性的。网络的权就可由线性方程组直接解出，从而大大加快学习速度并避免局部极小问题。

---

3 Matlab代码实现

GA.m

clear all
close all

G = 15;
Size = 30;
CodeL = 10;

for i = 1:3
    MinX(i) = 0.1*ones(1);
    MaxX(i) = 3*ones(1);
end
for i = 4:1:9
    MinX(i) = -3*ones(1);
    MaxX(i) = 3*ones(1);
end
for i = 10:1:12
    MinX(i) = -ones(1);
    MaxX(i) = ones(1);
end

E = round(rand(Size,12*CodeL));  %Initial Code!

BsJ = 0;

for kg = 1:1:G
    time(kg) = kg
    
    for s = 1:1:Size
        m = E(s,:);
        
        for j = 1:1:12
            y(j) = 0;
            
            mj = m((j-1)*CodeL + 1:1:j*CodeL);
            for i = 1:1:CodeL
                y(j) = y(j) + mj(i)*2^(i-1);
            end
            f(s,j) = (MaxX(j) - MinX(j))*y(j)/1023 + MinX(j);
        end
        
        % ************Step 1:Evaluate BestJ *******************
        p = f(s,:);
        
        [p,BsJ] = RBF(p,BsJ);
        
        BsJi(s) = BsJ;
    end
    
    [OderJi,IndexJi] = sort(BsJi);
    BestJ(kg) = OderJi(1);
    BJ = BestJ(kg);
    Ji = BsJi+1e-10;
    
    fi = 1./Ji;
    [Oderfi,Indexfi] = sort(fi);
    Bestfi = Oderfi(Size);
    BestS = E(Indexfi(Size),:);
    
    % ***************Step 2:Select and Reproduct Operation*********
    fi_sum = sum(fi);
    fi_Size = (Oderfi/fi_sum)*Size;
    
    fi_S = floor(fi_Size);
    
    kk = 1;
    for i = 1:1:Size
        for j = 1:1:fi_S(i)
            TempE(kk,:) = E(Indexfi(i),:);
            kk = kk + 1;
        end
    end
    
    % ****************Step 3:Crossover Operation*******************
    pc = 0.60;
    n = ceil(20*rand);
    for i = 1:2:(Size - 1)
        temp = rand;
        if pc>temp
            for j = n:1:20
                TempE(i,j) = E(i+1,j);
                TempE(i+1,j) = E(i,j);
            end
        end
    end
        TempE(Size,:) = BestS;
        E = TempE;
        
     %*****************Step 4:Mutation Operation*********************
     pm = 0.001 - [1:1:Size]*(0.001)/Size;
     for i = 1:1:Size
         for j = 1:1:12*CodeL
             temp = rand;
             if pm>temp
                 if TempE(i,j) == 0
                     TempE(i,j) = 1;
                 else
                     TempE(i,j) = 0;
                 end
             end
         end
     end
     
     %Guarantee TempE(Size,:) belong to the best individual
     TempE(Size,:) = BestS;
     E = TempE;
     %********************************************************************
 end

 
 Bestfi
 BestS
 fi
 Best_J = BestJ(G)
 figure(1);
 plot(time,BestJ);
 xlabel('Times');ylabel('BestJ');
 save pfile p;

RBF.m 

function [p,BsJ] = RBF(p,BsJ)

ts = 0.001;

alfa = 0.05;
xite = 0.85;
x = [0,0]';
b = [p(1);p(2);p(3)];
c = [p(4) p(5) p(6);
    p(7) p(8) p(9)];
w = [p(10);p(11);p(12)];

w_1 = w;w_2 = w_1;
c_1 = c;c_2 = c_1;
b_1 = b;b_2 = b_1;
y_1 = 0;

for k = 1:500
    timef(k) = k*ts;
    
    u(k) = sin(5*2*pi*k*ts);
    
    y(k) = u(k)^3 + y_1/(1 + y_1^2);
    
    x(1) = u(k);
    x(2) = y(k);
    for j = 1:1:3
        h(j) = exp(-norm(x - c(:,j))^2/(2*b(j)*b(j)));
    end
    ym(k) = w_1'*h';
    
    e(k) = y(k) - ym(k);
    
    d_w = 0*w;d_b = 0*b;d_c = 0*c;
    for j = 1:1:3
        d_w(j) = xite*e(k)*h(j);
        d_b(j) = xite*e(k)*w(j)*h(j)*(b(j)^-3)*norm(x-c(:,j))^2;
        for i = 1:1:2
            d_c(i,j) = xite*e(k)*w(j)*h(j)*(x(i)-c(i,j))*(b(j)^-2);
        end
    end
    
    w = w_1 + d_w + alfa*(w_1 - w_2);
    b = b_1 + d_b + alfa*(b_1 - b_2);
    c = c_1 + d_c + alfa*(c_1 - c_2);
    
    y_1 = y(k);
    w_2 = w_1;
    w_1 = w;
    
    c_2 = c_1;
    c_1 = c;
    
    b_2 = b_1;
    b_1 = b;
end

B = 0;
for i = 1:500
    Ji(i) = abs(e(i));
    B = B + 100*Ji(i);
end

BsJ = B;
        

Test.m

clear all;
close all;

load pfile;
alfa = 0.05;
xite = 0.85;
x = [0,0]';

%M为1时
M = 2;
if M == 1
    b = [p(1);p(2);p(3)];
    c = [p(4) p(5) p(6);
         p(7) p(8) p(9)];
    w = [p(10);p(11);p(12)];
elseif M == 2
    b = 3*rand(3,1);
    c = 3*rands(2,3);
    w = rands(3,1);
end

w_1 = w;w_2 = w_1;
c_1 = c;c_2 = c_1;
b_1 = b;b_2 = b_1;

y_1 = 0;

ts = 0.001;
for k = 1:1500
    time(k) = k*ts;
    
    u(k) = sin(5*2*pi*k*ts);
    y(k) = u(k)^3 + y_1/(1 + y_1^2);
    
    x(1) = u(k);
    x(2) = y(k);
    
    for j = 1:3
        h(j) = exp(-norm(x-c(:,j))^2/(2*b(j)*b(j)));
    end
    
    ym(k) = w_1'*h';
    e(k) = y(k) - ym(k);
    
    d_w = 0*w;d_b = 0*b;d_c=0*c;
    for j = 1:1:3
        d_w(j) = xite*e(k)*h(j);
        d_b(j) = xite*e(k)*w(j)*h(j)*(b(j)^-3)*norm(x-c(:,j))^2;
        for i = 1:1:2
            d_c(i,j) = xite*e(k)*w(j)*h(j)*(x(i) - c(i,j))*(b(j)^-2);
        end
    end
    
    w = w_1 + d_w + alfa*(w_1 - w_2);
    b = b_1 + d_b + alfa*(b_1 - b_2);
    c = c_1 + d_c + alfa*(c_1 - c_2);
    
    y_1 = y(k);
    w_2 = w_1;
    w_1 = w;
    c_2 = c_1;
    c_1 = c;
    b_2 = b;
    
end
figure(1);
plot(time,ym,'r',time,y,'b');
xlabel('times(s)');ylabel('y and ym');

pfile.mat 

p: [2.9915 2.9008 2.4982 1.0059 1.1056 0.8006 0.4780 1.6100 -1.3460 -0.7204 0.4076 0.2786]

4 结果