I need to design an appropriate Neural Network for my Data

Hi All

I am in need of correction of my neural network to work for my input and target data , please run the files with the NN , and see how the regression is

shall you please help me to get good results ?

[Greg Heath]

 currently the problem is that the input data are so similar and close to each other , sometime they differ from each other only in fraction , so the mapstd didn't really work

That doesn't make any sense.

Greg

[farzad]

Dear Professor

I do really need to find a solution , is there any ?

[farzad]

If you check my input file , you will see what I mean , please be kind to me and review the files , I will really appreciate it

[farzad]

I wish someone is reading the comments, I don't get a notification

[farzad]

Anyone on NN ?

[Greg Heath]

1. Did you obtain the following 19 plots?

   x(i=1:4,:) , t(j=1:3) and t vs x
   Anything remarkable?

2. What range of hidden nodes did you try?

   h= Hmin:dH:Hmax  = ?

3. How many random weight and data division trials for each value of h ?

   Ntrials = ?

4. Please tabulate your coefficient of determination results so I can compare with mine.

Greg

[Greg Heath]

Finally, plot the results figure, plot(result(:,1),result(:,2))

Hope this helps.

Greg

[farzad]

Thank you dear professor

well no I did not obtain the 19 plots , cause I don't understand it and don't know how it would help me , the hidden nodes I tried 10 .

my problem is that to me , the regression plot , bellow , does not look satisfying , and the solution is not precise , did it work the same for you ?

also how the plotting of the results will help me ???

and this is the tabulate of the coeff of determination :

result =

         0    0.7078
    2.0000    0.9463
    4.0000    0.9482
    6.0000    0.9458
    8.0000    0.9409
   10.0000    0.9443
   12.0000    0.9449
   14.0000    0.9485
   16.0000    0.9389
   18.0000    0.9492
   20.0000    0.9504
   22.0000    0.9440
   24.0000    0.9371
   26.0000    0.9401
   28.0000    0.9457
   30.0000    0.9424
   32.0000    0.9422
   34.0000    0.9484
   36.0000    0.9428
   38.0000    0.9401
   40.0000    0.9473

and it's plot is :

but for me , when I have to use the saved net , the results are not so satisfying

[Greg Heath]

1. Plot all 19. Then you tell me why you should have done this before designing the regression models.

2. The number of training equations are greater than the number of unknowns when

 H < Hub = 325

So you just choose Hmax = 40 ???

[farzad]

so 40 is wrong ? what number should I use ?

Answer 1

[Greg Heath]

 % result = 0      0.70784
%             3      0.91287
%             6      0.93873
%             9      0.94475
%            12      0.94233
%            15      0.94359
%            18      0.95516
%            21      0.95128
%            24      0.95247
%            27      0.94899
%            30      0.94737
% Elapsed time is 69.4 seconds.
% result =     0      0.70784
%                16      0.94645
%                32      0.95738
%                48      0.94561
%                64      0.95157
%                80      0.95522
%                96      0.94683
%               112      0.95382
%               128      0.96052
%               144      0.94832
%               160      0.96098
% Elapsed time is 265.5 seconds.

[farzad]

Thank you dear professor , so you mean that I should select Hmax = 160 ?

[farzad]

shall you please make it a bit more clear for me ? I don't understand

[Greg Heath]

Obviously, it takes at least 2 hidden nodes to approximate a single local max.

If you had plotted the outputs you would have seen that there a lot of local maxes in target 3.

Have you overlaid plots of output(red) on target plots(blue)?

Originally you indicated that 0.94 was an unacceptable result.

I showed that H = 160 ~ Hmax/2 will get you up to 0.96.

It is obvious what to do if you want to go higher with this topology.

However, if you want to exceed Hmax, then use a validation set or trainbr.

Another possibility is to use 2 or 3 separate nets

Hope this helps.

Greg

Answer 2

[Greg Heath]

My code that you included works ok. But I did have several comments

 1. PLOT ALL 19 COMPONENT PLOTS
 2. Plot results
 3. COMMENT OR DELETE the statemets

 inp=[ input(1,:) ]';
 netback

Hope this helps.

Thank you for formally accepting my anser

Greg