LINPROG requires the following inputs to be of data type double: 'f'.

I am trying to run an LP optimization.

My objective function is to estimate that minimize

where,

,

;

please note , in the code my reg1 , ...etc

I wrote the following code :

y= readtable('y.crra.csv'); % 249 x 1
y=table2array(y);
reg1=readtable('fitted.reg1.csv'); 
reg1=table2array(reg1); 
trareg1=reg1.';  % 1 x 249
reg2=readtable('u.fitted.reg2.csv'); 
reg2=table2array(reg2);
trareg2=reg2.';   % 1 x 249
reg3=readtable('u.fitted.reg3.csv'); 
reg3=table2array(reg3);
trareg3=reg3.';  % 1 x 249
reg4=readtable('u.fitted.reg4.csv'); 
reg4=table2array(reg4);
trareg4=reg4.';  % 1 x 249
tau=.05;
yp= @(x) x*trareg1 + x*trareg2+x*trareg3 + x* trareg4;
e= @ (x) y-yp;
loss= @ (x) max(tau*(e), ( tau-1 )*(e));
obj = @(x) sum(loss);
Aeq = [1, 1, 1, 1];
lb = [0, 0, 0, 0];
beq = [1];
x = linprog(obj, [], [], Aeq, beq, lb, []);

I received the following error :LINPROG requires the following inputs to be of data type double: 'f'.

do you have any idea how to solve it?

I searched about the error and I've seen other people mentioning this error but I don't get how could work in my case

Thanks in advance.

Answer 1

[Torsten]

y= readtable('y.crra.csv'); % 249 x 1
y=table2array(y);
reg1=readtable('fitted.reg1.csv'); 
reg1=table2array(reg1); 
trareg1=reg1.';  % 1 x 249
reg2=readtable('u.fitted.reg2.csv'); 
reg2=table2array(reg2);
trareg2=reg2.';   % 1 x 249
reg3=readtable('u.fitted.reg3.csv'); 
reg3=table2array(reg3);
trareg3=reg3.';  % 1 x 249
reg4=readtable('u.fitted.reg4.csv'); 
reg4=table2array(reg4);
trareg4=reg4.';  % 1 x 249
tau=.05;
loss = @(x) (x(1)*trareg1 + x(2)*trareg2+x(3)*trareg3 + x(4)* trareg4).' - y;
obj = @(x) sum(max([tau*loss(x),(1-tau)*loss(x)],[],2));
Aeq = [1 1 1 1];
beq = 1;
lb = [0 0 0 0];
ub = [1 1 1 1];
sol = fmincon(obj,0.25*ones(4,1),[],[],Aeq,beq,lb,ub)

[Az.Sa]

Thank you very much, that was very helpful as usual!

Is there any way to remove the the lower bound and upper bound but the sum of a_i =1 ? once I removed them the sum of the solution if not equal to 1

[Torsten]

But it's necessary to set at least

lb = [0 0 0 0];

isn't it ? Or would you also accept solutions x(i) < 0 ?

[Az.Sa]

I will accept negative x_i

[Torsten]

And using

y= readtable('y.crra.csv'); % 249 x 1
y=table2array(y);
reg1=readtable('fitted.reg1.csv'); 
reg1=table2array(reg1); 
trareg1=reg1.';  % 1 x 249
reg2=readtable('u.fitted.reg2.csv'); 
reg2=table2array(reg2);
trareg2=reg2.';   % 1 x 249
reg3=readtable('u.fitted.reg3.csv'); 
reg3=table2array(reg3);
trareg3=reg3.';  % 1 x 249
reg4=readtable('u.fitted.reg4.csv'); 
reg4=table2array(reg4);
trareg4=reg4.';  % 1 x 249
tau=.05;
loss = @(x) (x(1)*trareg1 + x(2)*trareg2+x(3)*trareg3 + x(4)* trareg4).' - y;
obj = @(x) sum(max([tau*loss(x),(1-tau)*loss(x)],[],2));
Aeq = [1 1 1 1];
beq = 1;
sol = fmincon(obj,0.25*ones(4,1),[],[],Aeq,beq)

gives a solution vector x that dosesn't sum to 1 ?

I just see there is an error in the code:

obj = @(x) sum(max([tau*loss(x),(tau-1)*loss(x)],[],2));

instead of

obj = @(x) sum(max([tau*loss(x),(1-tau)*loss(x)],[],2));

[Az.Sa]

yes , the results i received :

5.910446746214318e+14

1.283865325380095e+15

-1.435591081131757e+15

-4.393189188697679e+14

which is equal to 1.1250

[Torsten]

Did you correct the error before you ran the code ?

[Az.Sa]

Great !! Thank you very much for your help !!

Answer 2

[John D'Errico]

Apparently some or all of the arguments to LINPROG were not doubles. Which ones?

In your code, we see this:

yp= @(x) x*trareg1 + x*trareg2+x*trareg3 + x* trareg4;
e= @ (x) y-yp;
loss= @ (x) max(tau*(e), ( tau-1 )*(e));
obj = @(x) sum(loss);

So, is obj a double precision vector of numbers? (NO.) In fact, obj will return a SCALAR function of the argument vector.

obj is a function handle. While obj will return a number, if is NOT a vector of numbers. In fact, it is a NONLINEAR function of the argument, x. This is because of the max function inside the loss part of your objective.

Does LINPROG handle nonlinear objectives? (NO.)

Therefore, you cannot use LINPROG.

You may be able to use other tools like GA. But not FMINCON, because the max function inside the objective makes it not differentiable. And certainly not linprog.

[Az.Sa]

I am just trying to fit a quantile regression if you have heard about it. where the mean absolute deviation is special case of it when tau = .5 . But because I want to add the sum constraint I tried to use LINPROG

[John D'Errico]

(Yes, I know what a quantile regression is.)

But that does not mean you can use linprog by passing in a function handle. linprog does not accept functions as arguments. No place in the help for linprog does it suggest you can pass in a function handle for f. That is also true for intlinprog.

Try as hard as you wish, want what you like. But you cannot force code to do what you want instead of what it was written to do.

As I recall, you can write MAD problems using slack variables to be solved using linprog. I have done that in the past. But that means you need to reformulate the problem by adding in slack variables.

[Az.Sa]

can you please help with GA function code I have tried to edit my code based on that but it gives me error evry time

[John D'Errico]

Are you saying you want to solve the problem using GA? You would need to show the code you tried. But Torsten has already shown how to solve it using fmincon.