No Big Difference using newpr or patternnet and epochs reduction

I have used the following code , for my input and target file that I have attached as well , am I correct that newpr is Obsolete from MATLAB 2010 ? but when I replaced the newpr command with

patternnet(10)

there were some changes in number of the epochs , but not in the total results , I would like to ask your help to improve one of these codes to give the best results to the input that I have attached .

so I put the codes I have used here :

First I used this code :

close all, clear all, clc, plt = 0
load('input.txt')
%load input 
load ('target.txt')
%normalizing data 
input=input';
target=target';
x=input;
t=target;
% x = -2:0.1:2;
% t = sin(pi*x/2);
% [ x, t ]  = simpleclass_dataset;
[ I   N ] = size(x)   % [ 2 1000 ]
[ O N ] = size(t)    % [ 4 1000 ]
trueclass = vec2ind(t);  %vec2ind Transform vectors to indices.
class1 = find(trueclass==1);
class2 = find(trueclass==2);
class3 = find(trueclass==3);
class4 = find(trueclass==4);
N1 = length(class1)    % 243
N2 = length(class2)    % 247
N3 = length(class3)    % 233
N4 = length(class4)    % 277
x1 =  x(:,class1);
x2 =  x(:,class2);
x3 =  x(:,class3);
x4 =  x(:,class4);
plt = plt + 1
hold on
plot(x1(1,:),x1(2,:),'ko')
plot(x2(1,:),x2(2,:),'bo')
plot(x3(1,:),x3(2,:),'ro')
plot(x4(1,:),x4(2,:),'go')
 Hub = -1+ceil( (0.7*N*O-O)/(I+O+1))  % 399
Hmax = 40   % Hmax << Hub 
dH    = 4   % Design ~10 candidate nets
Hmin  = 2   % I know 0 and 1 are too small
 rng(0)   % Allows duplicating the rsults
j=0
for h=Hmin:dH:Hmax
    j = j+1
    net            = newpr(x,t,h);
    [ net tr y  ]  = train( net, x, t );
    assignedclass  =  vec2ind(y);
    err = assignedclass~=trueclass;
    Nerr           = sum(err);
    PctErr(j,1)    = 100*Nerr/N;
end
h = (Hmin:dH:Hmax)';    
PctErr = PctErr;
results = [ h PctErr ]

Then I changed from newpr to patternnet in the same code :

close all, clear all, clc, plt = 0
load('input.txt')
%load input 
load ('target.txt')
%normalizing data 
input=input';
target=target';
x=input;
t=target;
% x = -2:0.1:2;
% t = sin(pi*x/2);
% [ x, t ]  = simpleclass_dataset;
[ I   N ] = size(x)   % [ 2 1000 ]
[ O N ] = size(t)    % [ 4 1000 ]
trueclass = vec2ind(t);  %vec2ind Transform vectors to indices.
class1 = find(trueclass==1);
class2 = find(trueclass==2);
class3 = find(trueclass==3);
class4 = find(trueclass==4);
N1 = length(class1)    % 243
N2 = length(class2)    % 247
N3 = length(class3)    % 233
N4 = length(class4)    % 277
x1 =  x(:,class1);
x2 =  x(:,class2);
x3 =  x(:,class3);
x4 =  x(:,class4);
plt = plt + 1
hold on
plot(x1(1,:),x1(2,:),'ko')
plot(x2(1,:),x2(2,:),'bo')
plot(x3(1,:),x3(2,:),'ro')
plot(x4(1,:),x4(2,:),'go')
 Hub = -1+ceil( (0.7*N*O-O)/(I+O+1))  % 399
Hmax = 40   % Hmax << Hub 
dH    = 4   % Design ~10 candidate nets
Hmin  = 2   % I know 0 and 1 are too small
 rng(0)   % Allows duplicating the rsults
j=0
for h=Hmin:dH:Hmax
    j = j+1
    net            = patternnet(10);
    [ net tr y  ]  = train( net, x, t );
    assignedclass  =  vec2ind(y);
    err = assignedclass~=trueclass;
    Nerr           = sum(err);
    PctErr(j,1)    = 100*Nerr/N;
end
h = (Hmin:dH:Hmax)';    
PctErr = PctErr;
results = [ h PctErr ]

2 - Then I also tried to use another code that I found from this post : here

but I was not able to catch all the corrections that were advised by Proff. Heath , so I only used the original code. well there is a difference in confusion , that in the first code , only one of the green squares show a number like 189 / 100% , and only in one of the classes : 2 , not all of the four classes , why is that ?

3 - I also tried to use the sine(x) function by defining x vector and sin(x) , is input and target , but the code did not accept due to having 4 classes.

4 - And another question is , where in the first code , the curve fitting is done ?

5- how much can I rely on

net = init(net);
net = train(net,houseInputs,houseTargets);

and how many times repeat of it after running the code ?

• As an update I used mapstd after getting the inputs , but then doubted I should use mapminmax or mapstd ?

As an update I used

z=sim(net,input);

but it just created a vector of negative numbers between 1 and -1 , don't know how to use it !?

Answer 1

[Greg Heath]

% No Big Difference using newpr or patternnet and epochs reduction % Asked by farzad on 21 Feb 2015 at 13:32 % Latest activity Edited by farzad about 9 hours ago % % I have used the following code , for my input and target file that I have % attached as well , am I correct that newpr is Obsolete from MATLAB % 2010 ? but when I replaced the newpr command with patternnet(10) % there were some changes in number of the epochs , but not in the total % results , I would like to ask your help to improve one of these codes to % give the best results to the input that I have attached

Since you have the current PATTERNNET, FORGEDABOUD DA obsolete NEWPR.

However, when designing NON-OBSOLETE multiple nets in a for loop, the weights have to be EXPLICITLY initialized before training using CONFIGURE or INIT.

% so I put the codes I have used here :

Hmmm, look familiar (;>) !

SNIP

%Then I changed from newpr to patternnet in the same code :

However, you did not initialize the weights at the top of the inner for loop.

% 2 - Then I also tried to use another code that I found from this post : % here but I was not able to catch all the corrections that were advised % by Proff. Heath , so I only used the original code. well there is a % difference in confusion , that in the first code , only one of the green % squares show a number like 189 / 100% , and only in one of the % classes : 2 , not all of the four classes , why is that ?

Not understandable.

% 3 - I also tried to use the sine(x) function by defining x vector and sin(x) , % is input and target , but the code did not accept due to having 4 classes.

Doesn't make sense: That is regression, NOT classification!

% 4 - And another question is , where in the first code , the curve fitting is % done ?

No curvefitting: it's classification, not regression!

% 5- how much can I rely on % net = init(net); % net = train(net,houseInputs,houseTargets); % and how many times repeat of it after running the code ?

I'd guess somewhere between 10 and 90 per cent, PROVIDED you have the correct range of hidden node values AND you design ~ 10 nets for each value of hidden nodes.

% As an update I used mapstd after getting the inputs , but then doubted I % should use mapminmax or mapstd ?.

The CURRENT nets automatically normalize and unnormalize using the default MAPMINMAX. However, I prefer to use ZSCORE (easier to use than MAPSTD) BEFORE design to modify and/or delete outliers. Then I am too lazy to override MAPMINMAX, so I keep it.

% As an update I used % z=sim(net,input); % % but it just created a vector of negative numbers between 1 and -1 , don't know % how to use it !?

Assuming your classification target matrix had {0,1} unit matrix columns, you have to unnormalize your output and then use vec2ind to obtain the output classes.

The CURRENT syntax replaces SIM with

z = net(input);

Hope this helps.

Thank you for formally accepting my answer

Greg

[farzad]

Thank you so much professor heath

Yes it did help, only a new question that I have, is that where does the formula of the Hub come from? and do I need to design an I H O net? I mean are the calculation of Ntrneq and nw useful here? Would they improve? the results

[Greg Heath]

Number of training equations , Ntrneq, is not smaller than the number of unknown weights.

What topology do you think your net is???

Have you thought about searching the term Hub in the NEWSGROUP and ANSWERS?

Adding the additional search words Ntrneq and Nw might help.

[Greg Heath]

Hmm, just noticed that in the 1st case

net = newpr(x,t,h);

but in the second

net = patternnet(10);% NOT h!