This repository has been archived by the owner on Nov 19, 2020. It is now read-only.
Nelder-mead solver not converged #335
Comments
|
Hi there, I know it might be a bit late, but do you still have an example of the code that doesn't converge, so I could work on the issue? Regards, |
|
Please, check again this issue, as I am facing this problem as well. If you combine 2 parabolas, x^2-2*x and y^2+4*y to create some kind of valley, you get something like this:
Now, I start my NelderMead on x=0.5, y=0 (you may encounter the issue with other initial value as well). class Program
{
public static Double eval(Double[] val)
{
// WxMaxima command: plot3d(y^2+4*y+x^2-2*x,[x,-3,5], [y,-5,3],[grid,8,8]);
Double x = val[0];
Double y = val[1];
Double ret = y * y + 4 * y + x * x - 2 * x;
Debug.WriteLine("{2}; x={0}; y={1}", x, y, ret);
return ret;
}
static void Main(string[] args)
{
// This values are relevant for my RealWorld(TM) scenario
Double[] init = new double[] { 0.5, 0 };
NelderMead nm = new NelderMead(2, eval);
nm.Minimize(init);
// Solution
Debug.WriteLine("Minimum:");
Debug.WriteLine("{2}; x={0}; y={1}", nm.Solution[0], nm.Solution[1], nm.Value);
}
}And i get this output: Perhaps I am doing something wrong? |
|
Tried another example. I don't really understand what the algorithm is trying to accomplish. public static Double eval2(Double[] val)
{
// plot3d(y^2+4*y+x^2-2*x+(2*x^2-18*x)/(6*x^2-50*x+109)*(-y^2+5*y)/(2*y^2-8*y+9) ,[x,-4,6], [y,-6,4],[grid,80,80]);
Double x = val[0];
Double y = val[1];
Double ret = y * y + 4 * y + x * x - 2 * x + (2 * x * x - 18 * x) / (6 * x * x - 50 * x + 109) * (-y * y + 5 * y) / (2 * y * y - 8 * y + 9);
Debug.WriteLine("{2}; x={0}; y={1}", x, y, ret);
return ret;
}
static void Main(string[] args)
{
// Used this because (0,0) does nothing.
Double[] init = new double[] { 1, 1 };
NelderMead nm = new NelderMead(2, eval2);
nm.Minimize(init);
// Solution
Debug.WriteLine("Minimum:");
Debug.WriteLine("{2}; x={0}; y={1}", nm.Solution[0], nm.Solution[1], nm.Value);
}Output By the way, if you start in (0,0) it does not even move. It does 2 Iterations on (0,0) and then finishes. |
|
In case it helps, I post here the results for the SolverFoundation implementation, starting at (0,0). First function: Second function, does not find the minimum but the local: |
|
Hi @arubiomoreno, Thanks a lot for the informative examples. After some initial investigation, I could find at least one issue with the current code, in which the initial values of certain class variables are not being properly set and are causing the algorithm to terminate early after very few iterations. This will likely be fixed in a next release. Regards, |
cesarsouza
added a commit
that referenced
this issue
Jul 14, 2017
|
Many thanks for this. NelderMead nm = new NelderMead(...);
nm.GetType().GetField("minf_max", System.Reflection.BindingFlags.NonPublic | System.Reflection.BindingFlags.Instance).SetValue(nm, Double.NegativeInfinity); // TODO: Remove when 3.7.0 is releasedI post it here in case it is useful for someone |
|
Fixed in v3.7.0. |
Sign up for free
to subscribe to this conversation on GitHub.
Already have an account?
Sign in.
Hi, I tried to use the Nelder-mead solver for a customized optimization. I found with some initial inputs, the solver provided by accord.net can not converged. Actually the solver stops immediately after the 2nd iteration. However, if I feed the same initial input to the Nelder-mead solver provided by Microsoft Solver Foundation 3.1, the optimization can be carried on and get converged without any problem. Does anyone know how to enable the Nelder-mead solver in accord.net work in this case?
The text was updated successfully, but these errors were encountered: