Conjugate Gradient Methods Aren't Working

[Kreizhn]

I'm working on a control theoretical problem and trying to implement the solution in Matlab. Part of the solution requires minimizing a function f(x), for which my predecessor has opted to use a conjugate gradient method. He wrote his own conjugate gradient method, but it's not converging. I've replaced his method with a canned algorithm, but it is still not converging. This suggests to me that the problem is ill-suited to gradient methods.

Can anybody suggest to me why this might be the case? Is it likely because the surface lacks a sufficient degree of differentiability? Also, can anybody suggest another minimization algorithm that I could attempt to use instead of gradient descent?

 

[hotvette]

Hard to say without knowing more about the function. Can you tell us what it is?

You could try gradient descent method, but it may not converge any better.

http://en.wikipedia.org/wiki/Gradient_descent

Could also try Nelder-Mead Simplex:

http://en.wikipedia.org/wiki/Nelder–Mead_method

Random (Monte Carlo) method may work better if differentiability is an issue.