lsqcurvefit bad fitting results

Hello to everyone! I'm tryng to solve this data fitting problem with the following code, but as you see from the plot it seems doesn't fit well the points, what problem could be?

s = linspace(245,210,40);  % spring lenght
q =linspace(310*pi/180,340*pi/180,40); % swingarm angle
fun = @(x,s)(2*pi - (acos((x(8)^2+x(10)^2-sqrt(x(5)^2+x(6)^2-2*x(6)*x(5).*cos(acos((x(1)^2+x(2)^2-s.^2)/(2*x(1)*x(2)))+x(3)+x(4))).^2)/(2*x(8)*x(10))) - x(9) - x(7)) - x(7))*180/pi;
x0 = [117,222,31.4*pi/180,7.3*pi/180,45,210,30*pi/180,90,75*pi/180,185]; % initial guess
lb = [105,215,27*pi/180,5*pi/180,35,200,22*pi/180,75,65*pi/180,175] % Lower bound
ub = [125,235,37*pi/180,10*pi/180,55,225,37*pi/180,100,80*pi/180,195] % Upper bound
x = lsqcurvefit(fun,x0,s,q,lb,ub)
plot(s,q*180/pi,'ko') % data to fit
hold on
plot(s,fun(x,s),'b-') % optimized parameters
legend('data to fit','optimized parameters')
ylabel('swingarm angle (q)')
xlabel('spring lenght (s)')

This is the plot:

[Matt J]

The relationship between s and q is clearly linear. Not sure why you would be applying such a complex model function to something whose true model is far simpler (and is already known).

[Gregory Cottone]

I need to find the best parameters that can generate a linear function between s and q, imagine that I have a mechanism (posterior suspension of motorcycle) with many components with their relative geometric parameters, ok I need to optimize them. Let me know if I'm not clear.

I also would like to know if it is possible to make a multiobjective optimization with lsqcurvefit and other type of optimizations.

Thank you so much for the answer.

Answer 1

[Matt J]

Probably the sqrt() operation in your model function. There is nothing preventing the algorithm from searching in regions where the argument to the sqrt() is negative, resulting in complex numbers. The presence of sqrt() also means your model function will be non-differentiable, which violates assumptions of lsqcurvefit.

Answer 2

[John D'Errico]

Honestly, I would NOT solve it like that. Not even remotely so. Using lsqcurvefit for this just seems fraught with problems.

What might I do? I would look at the model, in the form of a Taylor series expansion, for the expansion of this function of s. (I'd also learn to use radians. Degrees are a bad thing, and what you have here is a frustrating amalgam of the two. It makes your code far more complicated than it need be.) Expand it around the midpoint of the interval of interest, thus around s at an extension of 225.

So your model wants to be in the form of a known extension at s = 225, plus a linear term, with a known slope.

Now I would look to find parameters that produce the desired linear behavior, while minimizing the higher order terms in that series expansion.

You know the relationship between s and q that you wish to find. It is just a straight line between the two end points. So now, look for parameters that kill off the higher order terms in the Taylor expansion, but maintain the desired low order terms as you need them.