Robust GA

Is their any way to test my GA method that it is not finding a local minima and that it is robust enough?

Many thanks in advance

Answer 1

[Walter Roberson]

Give it a test case such that one particular (multidimensional) point is associated with the value one, and every other (multidimensional) point is associated with the value 0. For example, the 3D case

(x(1)==pi && x(2) == exp(1) && x(3) == sqrt(2))

Or if you want to simulate a distinct local minima, you could use

(x(1)==pi) + (x(2) == exp(1)) + (x(3) == sqrt(2))

It would be interesting to see a GA that would, without prompting, find even one of the local minima, let alone the maxima.

[Hazem]

Thanks Walter,

I had my GA case and getting good results compared to a linear searching method. I want to proof in my thesis that it is not a local minimum point and that the method is robust

Any ideas please

Many thanks

Hazem

[Walter Roberson]

Without seeing your GA code, all I can suppose is that it is like the vast majority of other GA code, being confined to finding the global min or global max *probabilistically* rather than robustly.

Take a sample case and make the global maximum as narrow (improbable to wander in to) as is consistent with any domain constraints you might have. [Ideally, your code would have checks that would reject data that violated the constraints.] Is your GA code able to find the global max "robustly" under those conditions?