mlp.m

我ANN的大作业

function [u1,u2, v1, v2, b1, b2, times] = MLP(eps, initU1, initV1, initb1, initU2, initV2, initb2, alpha, input_P, target_T)
       
u1 = initU1;
v1 = initV1;
b1 = initb1;
u2 = initU2;
v2 = initV2;
b2 = initb2;

[len, dim] = size(input_P);
times=0;
index=1;
correct = 0;
lastcorrect=0;
while(true)
    x0 = input_P(index,:)';

    %先计算x1,x2,x3
    x1 = sigmoid(u1*(x0.^2) + v1*x0 + b1);
    x2 = sigmoid(u2*(x1.^2) + v2*x1 + b2);   
    t = target_T(index,:)';
    
    %这里可能还有问题
    result = sum((x2-t).^2);
    if(result<eps)
        correct = correct+1;
    else
        if (correct>=lastcorrect )
            correct
            times
            lastcorrect=correct;
        end
        correct = 0;
    end
    if(correct>=len)
        break;
    end
    
    %再计算s3,s2,s1
    s2 = -2*difF(x2)*(t-x2);
    s1 = difF(x1)*(2*u2*diag(x1)+v2)'*s2;
    
    %再迭代更新权值和偏置值
    u1 = u1-alpha*s1*(x0.^2)';
    v1 = v1-alpha*s1*x0';
    b1 = b1-alpha*s1;
    
    u2 = u2-alpha*s2*(x1.^2)';
    v2 = v2-alpha*s2*x1';
    b2 = b2-alpha*s2;
    
    times = times+1;
    index = mod(index+1,len+1);
    if(index==0)
        index=1;
    end    
end
len
times = times-len;

function y = difF(x)
y=diag((1-x)'*diag(x));