Model Fitting and Regression in MATLAB - YouTube

I'll be showing you some simple model fitting and regression techniques in MATLAB in this

screencast. What I am first going to do is show you how you can import some data from

Excel into MATLAB, and we're going to use that data in model fitting and regression

techniques in MATAB. I've got some data in a file called bacteria, and this is data

showing the optical density of a culture of bacteria as a function of time. You'll notice

for each of these times I have the intriplicate. Back in MATLAB I can go up here to the work

space and I can click "Import Data", and I will navigate to where my file is, I am going

to open that, and then you can see it's highlighted the data, it's not highlighting the title,

which is what I want, and so I can check this to make sure that that is what I want to import,

and then go up here to "Import" and you'll see that it imported this untitled, and I

can double click there to rename it, so I renamed this data. If you want you can actually

save this, so I am going to go up here to save, I'll click on that, and I can save this

to the folder that I want it to be in. At a later time point, if I want to reload that

data, I can just do load data, and assuming that I am in the right folder then it will

actually load it and it will but that array over here in the work space. Now what I am

going to do I am just going to plot this, so I am plotting data, all rows of column

1, and the response is all rows of column 2, and I am plotting these points as black

squares, and putting the grid on. So this is our plot, I've got time on the x-axis, optical

density on the left, and you can see that for each of these measurement times I've got

three different measurements of the bacteria culture. So what I am ultimately trying to

do is develop a model equation that models the growth of these bacterial cells. One way

to do this is to fit all those different data point to different polynomials, first defining

x as our time, y as the response, and here I've got p(1), it's a vector, which is the polyfit

of our data, a first-order polynomial, so just a line, I am creating an actual function

that's taking those coefficients of our fit and putting it into polyval, so this would

actually be a line so that we can plot that. I am doing second-order polynomial, third-order

polynomial, and then this fourth one is actually a logarithmic fit, so to do this using polyfit

you actually just take the polyfit of the log of y, and that will give us the equation

of a logarithmic fit. I am then computing the residuals, I have residuals one, which

is the residuals between our first-order polynomial, or our line and the actual values, I have

the second-order, third, and then I we have our logarithmic residuals. I am next plotting

a figure. I have subplots of all the different fits, so the first, second, and third-order

and then the logarithmic, but now I am going to run this file, so when I type bacteria,

that's the name of that m script that I showed all the plotting commands in, when I run this

it spits out two figures. The first one is a figure of all the model fits, so I have

the first-order, that's the line, second-order, third-order, and then our logarithmic fit.

The second plot, or the residuals for each of the fits, again the residuals are the difference

between the model and our actual experimental data points, you can see it's a quadratic

function which means that that's not a very good fit. We have our second-order residuals,

those look pretty good, they are somewhat random, I'd say our second and third are pretty

similar, and then we have our logarithmic fit. Bacteria actually grow exponentially,

so logarithmic should be the best fit, but these cells are actually in something known

as the lag-phase, so this is one way that we can plot our different fits and then analyze

to see how good of a fit each of those models is. So now I am going to show you how to do

these model fits and regression in MATLAB. If you go up here to tools in the figure,

and go down here to basic fitting you can choose all sorts of different fits, so I can

choose a linear fit, and you see that that is placed onto the plot. I can put a quadratic,

a cubic in there. If you want you can plot the residuals, it's actually going to tell

me that I can't use the bar plot residuals if I have replicates so I am going to do scatter

plot, we look down here this is the residual plot for all our different fits, I am actually

going to remove my linear, so I have two fits here, the quadratic and the cubic. If I want

I can expand this by clicking on this right arrow, and I can get the coefficients for

these fits, so here I have the cubic fit, if I want to look at a quadratic fit I can

do that, and these are the equations, and on the top here it has the form of that equation.

If I want to put in here a cubic fit, and then I can click on the right arrow here,

it will actually allow you to interpolate between points, so if I wanted to calculate,

let's say a time of 150, based upon the cubic model, then I can evaluate. I can also plot

that if I want, and it goes back here onto the figure, and then if you want to you can

save all of those to the work space. With the data fitting tool that I just showed you,

you are somewhat limited to the types of models that you can use, you have to basically go

with some of the models that are already built it. Fortunately there is a curve fitting tool

that is built into MATLAB, and you can actually create custom models. So let's do this in

MATLAB, I've imported a viscosity Excel file, so this is the viscosity of a fluid as a function

of the temperature, so I am going to just import this, rename it viscosity, and then I've extracted

the first column and that's my x-values, the second column is my y-values, and now if I

open up this cftool, which is curve fitting tool, it brings up this window, and I can

do a lot of things here, I can bring in my x-data, so my y-data, and using this I can

go over here and I can change the different type of model, here I am going to do a polynomial,

so this is a first-degree, which is just a line. I can go over here, I can see the goodness

of fit, so a lot of statistical parameters here. We want to maximize adjusted R-squared,

so I can go here and look at two, and you'll also notice down here I have the different

parameters, I can go to three for example, and I can basically go through a bunch of

different models to see which one gives me the best adjusted R-squared. Now I've put

in a custom equation, and we see that created a pretty good model here just with these three

terms, the adjusted R-squared is about 0.9972. So you can play around with a bunch of these

to determine the best fit for your data. This curve fitting tool can also work with multilinear

models, so that means you have two regressor variables, such as x(1) and x(2), you just

have two independent regressors and one output, either y or z, can have a second-order polynomial.

Unfortunately the curve fitting tool just by itself isn't capable of these interaction

terms, which would be like an x(1) times an x(2), or an x times y, but you can easily

put in a custom tool into the curve fitting tool. So here I've got some multilinear data.

In column A I have the response, which is, we're going to say is z, and then I've got two regressor

variables, x and y, and I am going to import this, and I am going to rename, and then I

am extracting my z-values, that's the first column of our data, the x values is the second

column, and the y-values is the third column. Now if I type in cftool that brings up the

curve fitting tool, and I brought in my x, my y, and my z, and this time you can actually

play around with this three-dimensional plot to see how well that curve, and right now

I have a polynomial with a degree one for each of those, so the equation is over here

on the left. If I want I can name this fit and it will actually store it down here as

first -order, I can create a new bit, and I've adjusted the order of two for both x

and y, and I can save that and I can compare the two at the bottom. Finally if I want to

do the interaction, which was the third one I showed on that slide, I can go here to custom.

So I've created a new fit here, custom brought xyz, now I have axy squared, plus by squared,

plus cxy, so this cxy is the interaction multiplying those two that you can't do with a typical

fit, and I look down here, I look, I can rotate this and see that's not actually as good of

a fit, and looking down here at the adjusted R-squared, that's actually not as good of

a fit as the second-order. So the second-order, the one shown here, actually fits the data

very nicely.