# How to get function of "GaussianProcess" prediction?

I am a new user of Wolfram Mathematica. I need your help.

Let's say I have three variables: x, y, z. I want to calculate the "z" depending on x, and y. In other words, I want to find a function for z like z=f(x,y).

I have this data

y = {140, 360, 500, 740, 800};

x = {4, 10, 35, 70, 90};

z = {{97.438, 103.891, 110.344, 118.7545, 124.13475},
{110.344, 116.797, 123.25, 127.165, 129.515},
{118.7545, 122.95975, 127.165, 131.08, 131.865},
{121.444625, 124.8923125, 128.34, 131.4725, 132.2575},
{124.13475, 126.824875, 129.515, 131.865, 132.65}
};


based on which I have

1. Created the plots of the data ListPlot3D[vals, Mesh -> All]

2. Created the predict and prediction is by "LinearRegression"

trainingSet =
Flatten[Table[
Rule[{s[[i]], t[[j]]}, data[[i, j]]], {i, Length@s}, {j,
Length@t}], 1];
pf = Predict[trainingSet, Method -> "LinearRegression"];
Show[Plot3D[pf[{x, y}], {x, Min@s, Max@s}, {y, Min@t, Max@t}],
ListPointPlot3D[vals, PlotStyle -> {PointSize -> Large}]]
[![enter image description here][3]][3]


LinearRegression results.

it doesn't represent all the points that I need.

1. By "PredictorInformation[pf, "Function"]" I get the function of my data. The function is = "109.712 + 0.115408 #1 + 0.0168067 #2 &"

However, the prediction is linear which doesn't fully give me what I want to see. There is another prediction method which is "GaussianProcess" and it gives me what I want it to see.

but when I try to get the function of it by "PredictorInformation[pf, "Function"]"

there is an error "PredictorInformation::elmntavs: "Function" is not an available property. Did you mean "FunctionMemory" instead?"

So my question is: Is there a way of getting the function of GaussianProcess prediction?

Or if there is not a way for that. How can I get a function of z = f(x,y) that would represent my data properly?

• If I'm not mistaken, the full functional expression of a fitted Gaussian process would be a pretty big and ugly beast that basically involves all data points that were used to make the fit as well as a matrix inversion. A Gaussian process is more like an interpolation function than a linear model in that sense. That's probably the reason they don't give you the function explicitly. Commented Feb 16, 2019 at 12:06
• Thank you for your comment. Is there another way to get a function of z = f(x,y) that would represent my data properly? Commented Feb 16, 2019 at 12:17
• If you just want an analytical formula, you should probably use Fit or NonlinearModelFit (a 2nd degree model will probably do). Predict is a machine learning method that is geared towards prediction of unseen values and not so much towards fitting a model you can easily interpret. Commented Feb 16, 2019 at 12:40
• Is there a guide I can use or some examples of this kind of data with z = f(x,y)? Commented Feb 16, 2019 at 13:00

Here's how you can fit a 2D polynomial to your data:

y = {140, 360, 500, 740, 800};

x = {4, 10, 35, 70, 90};

z = {{97.438, 103.891, 110.344, 118.7545, 124.13475}, {110.344,
116.797, 123.25, 127.165, 129.515}, {118.7545, 122.95975, 127.165,
131.08, 131.865}, {121.444625, 124.8923125, 128.34, 131.4725,
132.2575}, {124.13475, 126.824875, 129.515, 131.865, 132.65}};
fitdata = Catenate @ Table[{x[[i]], y[[j]], z[[i, j]]}, {i, 1, Length[x]}, {j, 1, Length[y]}];
fit = Fit[fitdata, {1, x1, x2, x1 x2, x1^2, x2^2}, {x1, x2}]
Show[
Plot3D[fit, {x1, Min[x], Max[x]}, {x2, Min[y], Max[y]}, PlotStyle -> Blue],
ListPlot3D[fitdata]
]


Out[21]= 94.7557 + 0.645641 x1 - 0.00383606 x1^2 + 0.0364044 x2 - 0.000257039 x1 x2 - 2.21253*10^-6 x2^2

• Hi, thank you so much it worked for me. I changed the fit data equation to (1, x1, x2, x1^2, x1 x2, x2^2, x1^3, x1^2 x2, x1 x2^2, x2^3) and got good results. Also as you suggested I tried to use "NonlinearModelFit" which gave me the "BestFitParameters" but they don't seem to be the best fit parameters. Really thank you so much I appreciate your help. If didn't publish this example I would struggle all day long. Commented Feb 17, 2019 at 14:39
• With NonlinearModelFit, you pretty much always need to specify good initial guesses for your parameters. That's just the nature of fitting nonlinear models: there's no algorithm that guarantees a good result. Linear fits (i.e., linear in the basis functions) are much easier in that respect. Commented Feb 17, 2019 at 18:45
• I have also found the "FindFormula". Right now don't get how it works but I am trying to find out. Commented Feb 18, 2019 at 15:22
• FindFormula does not work for 2D data AFAIK. Commented Feb 18, 2019 at 15:40