fsolve in codegen returns indexing error

Hello,

I am struggling with the code generation of a function containing the fsolve command (supported since R2020b). Since it is a rather new topic, I did not find any related entries.

This is the function i try to generate the mex-file for:

function [x] = test_function(a,b)
%#codegen
x       = zeros(1);
opts    = optimoptions('fsolve','Algorithm','levenberg-marquardt','Display','off');
% opts    = optimoptions(opts,a.string,a.val);
fcn     = @(in)feval(b,in);
rng default
x0  = double(0.00);x(1)  = fsolve(fcn,x0,opts);
end

And this is how I setup the code generation:

cfg = coder.config('mex');
types = coder.getArgTypes('mytest','test_function')
codegen -config cfg -args types test_function 

With the following function to get the coder.Types:

function out =  mytest
a.val = 10e-5;
a.string = "FunctionTolerance";
b = 'myeq';
out = test_function(a,b);
end

The struct a has no purpose yet, but will later be used to setup the optimoptions.

Error Message after calling the codegen:

"??? Index expression out of bounds. Attempted to access element 1. The valid

range is 1-0."

In the Coder Report Viewer it sais, that the error occured in function projectBox. I would appreciate any thoughts on this.

Many Regards,

Max

[Max Bartholdt]

I have not really solved the issue. But changing the number of input parameters in myeq from 1 to 2 makes it possible to generate code.

Why does it not work if I am using just 1 input parameter in function myeq?

[Mukund Sankaran]

Hi Max. Would it be possible for you to share the code for myeq ?

[Max Bartholdt]

function x = myeq(u)
x = zeros(1,1);
u1 = u(1);
% u2 = u(2);
x(1) = (u1^2-u1)-3;
% x(2) = u2*u1+2;
end

Answer 1

[Caleb Magruder]

Hi Max,

Unfortunately, this is due to a bug which will be fixed in an upcoming software update.

In the meantime, you can work around the issue by calling lsqnonlin instead of fsolve and pass -Inf as a lower bound and Inf as an upper bound, effectively making the problem unbounded. You will also need to update the optimoptions solver from 'fsolve' to 'lsqnonlin' in line 4 of test_function. The updated test_function is below for your reference.

We apologize for the inconvenience this has caused and appreciate you bringing it to our attention.

function [x] = test_function(a,b)
%#codegen
x       = zeros(1);
opts    = optimoptions('lsqnonlin','Algorithm','levenberg-marquardt','Display','off');
% opts    = optimoptions(opts,a.string,a.val);
fcn     = @(in)feval(b,in);
rng default
x0  = double(0.00);x(1)  = lsqnonlin(fcn,x0,-Inf,Inf,opts);
end

[Max Bartholdt]

Hello Caleb,

Thank you for your quick answer and the suggestion for a workaround.

Many Regards,

Max