Wouter Kool

* I think the largest opportunity (also applying pre-checks with existing timewarp) has been implemented using the new cost evaluator functions. * My previous comment is no longer valid as...

Sorry, I have not been able to catch up lately but it looks like a lot of nice work which indeed replaces this PR. I will find some time beginning...

Closing this since it was superseded by the new structure which I confirmed works as expected.

I think all of the complexity in my previous solution was to incrementally maintain the biased fitness... I think I did not try just computing it lazily when needed using...

Indeed the exact SPLIT implementation is here: https://github.com/vidalt/HGS-CVRP/blob/main/Program/Split.cpp or here: https://github.com/ortec/euro-neurips-vrp-2022-quickstart/blob/main/baselines/hgs_vrptw/Split.cpp. I think this has `O(n)` complexity and is optimal so I don't see much benefit of implementing a heuristic...

@N-Wouda, do you have any specific thoughts on how to best handle this, e.g. use the appropriate distance function for the right problem variants?

I had some discussions with a student who investigated the parent selection for (undirected) CVRP and found a lot of duplicate solutions are in the population which aren't recognized and...

> I think we first have to figure out if having symmetric BPD matters or not. I suggest implementing the symmetric BPD, which should not be too hard, and then...

It's a small change so let's try; I've opened #613. Note this probably conflicts with #586 but that should be easy to resolve.

Summary of considerations/ideas: * Only based on centroid + tolerance (as implemented in #251) seems too limited to capture variety of routes. * Doing a circle sector that captures all...