How to increase the number of data in 70x6 matrix?

Dear All,

I wonder if MATLAB has a tool allows for increasing the number of data in 70x6 matrix while maintinaing the same relations. Some thing like intrapolation but for several arrays instade of one.

Thanks

Aziz

[Walter Roberson]

What would be the difference between that and interp2() ?

[Abdulaziz Abutunis]

Hi Walter,

I am not sure if interp2() can handel it because I am not very familiar with it, though I will look at it.

My experimental data are not sorted from hihgh to low or low to high. It was generated based on desing of experiment, so some inputs are fixed while others are vaired to inspect the responses behaviour. I am not sure if interp, in general, can understand/deal with this behaviour.

Thank you for your suggesstion. Again, I will check and let you know

Best

Aziz

[Walter Roberson]

You can use interp2() provided that you can:

• create (sorted) marginal vectors for the two independent variables, OR
• create a lookup matrix for each of the two independent variables, with the lookup matrices being monotonic in both directions (but not strictly monotonic) . For example [1 1 1; 2 2 2; 3 3 3] is fine as would be [1 1.01 1.02; 2 2.02 2.04; 3; 3.03; 3.06] .

This is enough flexibility to, for example, create longitude references to physical coordinates: the grid does not need to be equidistant, as long as it does not "cross".

[Abdulaziz Abutunis]

This looks smart idea, so thank you. However, my matrix has 5 inddependent veriables, do you think I still can use this method with interp2().

Regards

Aziz

[John D'Errico]

Um, what does a 70x6 matrix mean? What are the two dimensions of that matrix? Do you have 70 data points that essentially live scattered in a 6 dimensional spade? Or do you have a 70x6 array, that one can imagine as a 2-dimensional surface, thus z(x,y)? I suppose you might even have only 6 data points, each of which lives in a 70 dimensional space. The last case would be hopeless to solve of course.

The difference is highly significant, and would require different solutions for each case.

[Abdulaziz Abutunis]

As I stated earlier they are 5 independent variables and one dependent variable. So it is 5D not 2D. The 70 are the leverls of these variables, but repeated sometimes

Below is an example of 54 data inputs

Thanks

5  24  1  11  400 
5  24  1  11  800 
5  24  1  11  1200 
5  66  2  14  400 
5  66  2  14  800 
5  66  2  14  1200 
5  129  3  17  400 
5  129  3  17  800 
5  129  3  17  1200 
8  24  2  17  400 
8  24  2  17  800 
8  24  2  17  1200 
8  66  3  11  400 
8  66  3  11  800 
8  66  3  11  1200 
8  129  1  14  400 
8  129  1  14  800 
8  129  1  14  1200 
6  24  3  14  400 
6  24  3  14  800 
6  24  3  14  1200 
6  66  1  17  400 
6  66  1  17  800 
6  66  1  17  1200 
6  129  2  11  400 
6  129  2  11  800 
6  129  2  11  1200 
5  24  1  11  400 
5  24  1  11  800 
5  24  1  11  1200 
5  66  2  14  400 
5  66  2  14  800 
5  66  2  14  1200 
5  129  3  17  400 
5  129  3  17  800 
5  129  3  17  1200 
8  24  2  17  400 
8  24  2  17  800 
8  24  2  17  1200 
8  66  3  11  400 
8  66  3  11  800 
8  66  3  11  1200 
8  129  1  14  400 
8  129  1  14  800 
8  129  1  14  1200 
10  24  3  14  400 
10  24  3  14  800 
10  24  3  14  1200 
10  66  1  17  400 
10  66  1  17  800 
10  66  1  17  1200 
10  129  2  11  400 
10  129  2  11  800 
10  129  2  11  1200 

Answer 1

[Walter Roberson]

Ah, my earlier replies were based upon the idea that the data is 2D.

If you had one fewer independent variables then you could use scatteredInterpolant, but you cannot do that with 4 independent variables.

You can compute the voronoi diagram; https://www.mathworks.com/help/matlab/ref/voronoin.html but it is not immediately clear what you would do with the result.

It might be easiest to do a KNN (K-Nearest Neighbours) classification and use predict() at the points you want to fill in; https://www.mathworks.com/help/stats/fitcknn.html and https://www.mathworks.com/help/stats/fitcknn.html

[Abdulaziz Abutunis]

Thank you, Walter, for the suggestions

I am increasing the # of data because I want to use them to train artificial neural network. I will check you suggestion.

Regards

Aziz

[Walter Roberson]

You cannot meaningfully train on predicted or interpolated data for most purposes.

... Except for the case for where the reason you are training is to find a replacement algorithm for the prediction or interpolation, having already verified by some other means that the prediction or interpretation does a good job, but perhaps looking for a replacement algorithm that can be executed faster (such as on GPU)

[Abdulaziz Abutunis]

Agree, thanks again