Need help! I try to build Neural Network but it does not fit well with the training data!

Question

Quang Cung 2018-11-25

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/431781-need-help-i-try-to-build-neural-network-but-it-does-not-fit-well-with-the-training-data

评论： Quang Cung 2018-11-26

采纳的回答： Greg Heath

在 MATLAB Online 中打开

I check my feed forward and backpropagation but cant find error, please help! Thank you so much.

Here is my code:

%% Prepare data
% train data:                               
I=1;                                % number of independent variables
F=1;                                % number of dependent varaiables
Ib=[-pi:0.1:(4*pi)].';              % input matrix
Yb= sin(Ib);                        % (actual) output matrix
P=size(Ib,1);                       % number of observations
% Normalize:
Input(:,1)=(Ib(:,1)-min(Ib(:,1)))/(max(Ib(:,1))-min(Ib(:,1)));
Y(:,1)=(Yb(:,1)-min(Yb(:,1)))/(max(Yb(:,1))-min(Yb(:,1)));
%% Neural Network:
% hyperparameter for NN:
Gen=10^8;                   % number of times train neural network
Epsi= 10^(-7);              % tolerance
SSEold=999;                 % sum squared error
m=0;                        % update step size
A=1;
B=10000;
mu=A/(B+m);                 % step size or rate of learning
% layer:
N=1;                        % number of hidden layers
H=15;                       % number of hidden nodes
g=0.01;
% Initial parameter:
W1=rand(I+1,H);                 
X=g.*rand(H,F);
bias=g*rand(P,1);           % bias node
bias1=g*rand(1,1);          % bias weight to output layer
% Add bias node to Input layer:
Input=cat(2,Input,bias);
Err=10;
while (Err >Epsi && m < Gen)
%% Forward probagation:
    Vp=Input*W1;
    Vp=1./(1+exp(-Vp));        
    Yhat=bias1+Vp*X;
    SSEnew=sum(abs(Y-Yhat).^2,1);   
%% Backward probagation:
    Yerr=Y-Yhat;
    bias1=bias1+mu*sum(Yerr);
    W1=W1+mu*(Input.')*(Yerr*(X.').*Vp.*(1-Vp));
    X=X+mu*(Yerr'*Vp).';
%% Update:
    m=m+1;
    mu=A/(B+m);
%% Calculate Sum Squared Error:
    Err=abs(SSEnew-SSEold);
    SSEold=SSEnew;
end
%% plot
figure1=figure;
grid on
plot(Y)
hold on
plot(Yhat)

0 个评论
显示 -2更早的评论隐藏 -2更早的评论

请先登录，再进行评论。

请先登录，再回答此问题。

Answer 1

Greg Heath 2018-11-26

1
链接

此回答的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/431781-need-help-i-try-to-build-neural-network-but-it-does-not-fit-well-with-the-training-data#answer_348878

>> Vp = Input * W1;

Error using *

>> size(Input), size(W1)

ans =

158 4

ans =

2 15

Hope this helps

Thank you for formally accepting my answer

Greg

1 个评论
显示 -1更早的评论隐藏 -1更早的评论

Quang Cung 2018-11-26

在 MATLAB Online 中打开

Hi Greg Heath, that was a good catch, thanks for that, but I still got a problem with the code, the result from NN does not fit well with the train data set. Here is the code after I modified, please help me to fix it. Thanks.

clear all
clc
%% Prepare data
% =========train data===========                               
Ib=[-pi:0.1:(4*pi)].';              % input matrix
Yb= sin(Ib);                        % (actual) output matrix
% ==========Normalize============
Input(:,1)=(Ib(:,1)-min(Ib(:,1)))/(max(Ib(:,1))-min(Ib(:,1)));
Y(:,1)=(Yb(:,1)-min(Yb(:,1)))/(max(Yb(:,1))-min(Yb(:,1)));
%% Neural Network:
P=size(Input,1);            % number of observations
I=size(Input,2);            % number of independent variables
F=size(Y,2);                % number of dependent varaiables
% =========operational parameter for NN===========
Gen= 10^(6);                 % number of times train neural network
Epsi= 10^(-7);              % tolerance
m=0;                        % update step size
A=1;                        
B=10000;
mu=A/(B+m);                 % step size or rate of learning
SSEold=999;                 % sum squared error
% ============layer================================
N=1;                        % number of hidden layers
H=15;                       % number of hidden nodes
g=0.5;
% ============Initial parameter====================
W1=rand(I+1,H);                 
X=g.*rand(H,F);
bias=g*rand(P,1);           % bias node
bias1=g*rand(1,1);          % bias weight to output layer
% =============Add bias node to Input layer========
Input=cat(2,Input,bias);
Err=10;
while (Err >Epsi && m < Gen)
%% Forward probagation:   
    Vp=Input*W1;
    Vp=1./(1+exp(-Vp));        
    Yhat=bias1+Vp*X;
    SSEnew=sum(abs(Y-Yhat).^2,1);
%% Backward probagation:
    Yerr=Y-Yhat;
    bias1=bias1+mu*sum(Yerr);
    W1=W1+mu*(Input.')*(Yerr*(X.').*Vp.*(1-Vp));
    X=X+mu*(Yerr.'*Vp).';
%% Update:
    m=m+1;
    mu=A/(B+m);   
%% Calculate Sum Squared Error:
    Err=abs(SSEnew-SSEold);
    SSEold=SSEnew;
end
%% plot
figure1=figure;
grid on
plot(Y)
hold on
plot(Yhat)