Segmentation Fault EGO

Issue #11 new
mashall aryan
created an issue

I get Segmentation fault from running EGO on an 18 dimensional expensive optimisation problem. A set of ~320 or more observed points exist in the beginning. I have included the log file and the ones related to initial and final states.

Comments (2)

  1. Ruben Martinez repo owner

    If you recompile the library in DEBUG mode (for example: cmake -D CMAKE_BUILD_TYPE=Debug .), it will give more information about the Segmentation Fault.

    Meanwhile, It seems that your function is ill-defined in terms of the GP as the optimal parameters are all [1, 1, ... ,1]. It could be that the noise level is wrong (too high or too low), also the sGaussianProcessML is known to be unstable. Better use a standard GP or a Student's t process. SC_MAP is also more stable than SC_ML.

    The "hacking BOBYQA" message is because the optimal is in the boundary, that might also a signal that the model is wrong.

  2. Log in to comment