MATLAB: Solving 8-equations system in 8 variables with lsqnonlin()

Question

Hi I'm trying to use lsqnonlin with this system:

'file.m' if true % code

function y = prova_in_s(x,Sm,S_std,g_m,g_l)

T1=tras_scatter_par(Sm(:,1:2)); T_std1=tras_scatter_par(S_std(:,1:2)); S_probe=[x(1)+i*x(5),x(2)+i*x(6);x(3)+i*x(7),x(4)+i*x(8)]; T_probe=tras_scatter_par(S_probe);

parz=[real(T_probe*T_std1*T_probe-T1); imag(T_probe*T_std1*T_probe-T1)];

if numel(Sm)>4 T2=tras_scatter_par(Sm(:,3:4)); T_std2=tras_scatter_par(S_std(:,3:4));

parz2=[real(T_probe*T_std2*T_probe-T2); imag(T_probe*T_std2*T_probe-T2)];

% parz2=[real(T_probe*T_std2); % imag(T_probe*T_std2)]; %

end

for cicles=1:numel(g_m) loadst(cicles)=(g_l(cicles)*(T_probe(1,2))/(T_probe(1,1))+1)*g_m(cicles)-g_l(cicles)*det(T_probe)*(T_probe(2,2))/(T_probe(1,1))-((T_probe(2,1))/(T_probe(1,1))); end

for cicles=1:numel(g_m) if cicles==1 parz_load=[real(loadst(cicles));imag(loadst(cicles))]; else parz_load=[parz_load;real(loadst(cicles));imag(loadst(cicles))]; end end

y1=[parz(1);parz(2);parz(3);parz(4);parz(5);parz(6);parz(7);parz(8);parz_load];

y2=[parz(1);parz(2);parz(3);parz(4);parz(5);parz(6);parz(7);parz(8)];

if numel(Sm)>4 y=[y1;y2]; else y=y1;

end end

The problem consists in:

varyng the coefficients of the equations the solver calculate a wrong solution with an error of 20-30 % from the real solution. I thought that a reason is that the solution is not unique…so the choose of the intial guess, x0, is fundamental…i could adjust the FunTol too..anyway i call lsqnonlin in this form: " [x_sol,resnorm] = lsqnonlin(@(x)prova_in_s(x,a,b,c,gamma_std),x0,lbo,ubo); " how can i adjust FunTol? if I try to pass the option structure it return me an error. I read also thet optimset is not valid for lsqnonlin, so how i can adjust the options like algorithm etc.. Thank you in advance for your help.

Answer 1

optimset indeed is supported by LSQNONLIN. Just make sure its the 5th argument. Not all options are available for all solvers, but here is reference that can help:

http://www.mathworks.com/help/optim/ug/optimization-options-reference.html

From your objective function it seems like it is combinatorial in nature. I am not sure LSQNOLIN is perfectly suited for this type of problems, which is more suited for least squares (min sum of squares) kind of problems. Which are always smooth. But with jumps and discontinuities, I would rather go with a global optim approach such as GA which is stochastic and can handle such behavior.