resibots/limbo

the new sample is nan (not a number) or infinity

langongjin opened this issue · 3 comments

langongjin commented

hi, I got a problem that the new sample is nan (not a number) or infinity when I test Limbo for 2 dimensions function.

I printed the new sample ("error from bo_base.hpp : nan"). The error as shown following:
37 new point: 0.0123711 8.46754e-06 value: 0.999923 best:0.999982 38 new point: 0.0205997 0.0554797 value: 0.998247 best:0.999982 error from bo_base.hpp : nan libc++abi.dylib: terminating with uncaught exception of type limbo::EvaluationError: std::exception

This did not happen for one dimension objective function. sometimes it happens for two dimensions. However, it happened more often for three dimensions. furthermore, I always happen for four dimensions.

I know the new sample is checked by function "is_nan_or_inf ()" in tools/math.h
But why this problem happen for two dimensions function but not for one dimension, and how to solve this problem?

langongjin commented

hi, Does anybody has an idea about this question, why and how to solve? Thanks!

jbmouret commented
langongjin commented

hi, Thank you very much!

It is working due to the inspiration from your answers. The problem is that the evaluation function is not right in high dimensional space.

Thanks again!