Resample when a point is out-of-bounds

[msarehn]

Adds resampling to keep points within defined bounds. Compared to other constraint-handling methods, this can speed up convergence significantly. It's not a one-size-fits-all solution, however.

Also supports a resampling limit in order to not get stuck resampling on the boundary's edges.

[pengowen123]

This condition should be checked in CMAES::new, and can you add a test for it as well? Also, minor typo: defalt -> default.

[msarehn]

Hm, when I go the dyn Constraints way, this will not be possible anymore. We don't know what the concrete Constraints require. I could probably try to downcast do Bounds specifically, but that would be a code smell I think.

What do you think?

[pengowen123]

A clean way to verify constraints would be adding a method like fn are_constraints_valid(&CMAESOptions) -> bool to Constraints and calling it in CMAES::new. I'm fine with leaving this out for now because the code providing the constraints can ensure their validity.

[pengowen123]

Thanks for the PR! I left some minor comments, and can you also run rustfmt?

This feature is definitely useful, and the implementation looks good as well. I'm wondering though if it would be better to generalize Bounds to a trait like this:

trait Constraints {
    fn meets_constraints(x: &DVector<f64>) -> bool;
}

Sampler would store a Box<dyn Constraints>, and the existing Bounds can simply implement the trait. The trait enables support for more complex constraints at the cost of a virtual function call per point sampled. I imagine this cost would be small though, especially relative to expensive objective functions. What do you think?

[msarehn]

Went the dyn Constraints way now. Also ran rustfmt. Ready for re-review.

[pengowen123]

Looks great! I left one more comment, but I'll merge this as soon as it's resolved.

[msarehn]

Squashed the fixups.

[pengowen123]

Thanks!