Description
First off: thank you so much for skopt - I really like the API!
I was looking in the documentation for best practices advice when the user supplied callback throws an exception or returns NaN when using e.g. forest_minimize, but I couldn't find any advice. I hope it is alright that I open an issue regarding this?
Short background: I am trying to optimize parameters in a model. My callback calculates the RMSD between experimental data and model prediction. The model prediction data is calculated by integrating a system of ODEs (using e.g. scipy.integrate.ode). For a small set of parameters explored by e.g. forest_minimize the integration fails.
Currently I keep track of the biggest value seen by my callback and always return that value + some arbitrary number (e.g. 1) for sets of parameters for which my callback fails. This obviously distorts the cost-function landscape. Is there any advice to be given here? (except from making my callback robust for any input)