317 crowding distance calc needs to handle single individual pops #318
Conversation
…population of a single individual, in which case we should assign a distance of 'inf' to indicate it has no neighbors. We also return the population that was given since the sorting mechanism may yield empty lists, which is what we were previously returning.
…vidual populations
|
I thought individuals in ranks of size 1 would already have a distance of 0, since all objectives would be degenerate (range of 0) and I changed it to skip those objectives for setting infs. I just tested it, though, and numpy would give a nan value for an objective range if the rank has solely nan or +/- inf fitnesses in that objective, which is seen as non-degenerate since its non-zero. This is probably undesirable, even if those fitnesses usually denote some other issue, since crowding distances would be wrong. The fix is simple so I'll make a PR to this branch for it, but do you know if that was the case when Robert encountered the bug? Also, I see where the bug was with returning |
|
Can we merge this into Also could use unit tests (sorry for repeating this in two separate places). |
…needs-to-handle-single-individual-pops Also skip nan range objectives, which are degenerate if fitnesses are inf
…-to-handle-single-individual-pops' into 317-crowding-distance-calc-needs-to-handle-single-individual-pops
…e-single-individual-pops
… a single threaded worker. Hypothesis was that there were too many workers spun up by default that consumed too many resources. Besides, we only need a single worker given the nature of the distributed unit tests.
…-to-handle-single-individual-pops' into 317-crowding-distance-calc-needs-to-handle-single-individual-pops
|
The Windows python 3.9 check is failing for one of the distributed unit tests. It doesn't fail anywhere else, so there's something about that magic combo that's causing problems. At this point I don't have the bandwidth to track this down; if we have someone on Windows that encounters this problem, then we can take a deeper dive. However, I don't think we have any users that match that criteria. |
markcoletti
deleted the
317-crowding-distance-calc-needs-to-handle-single-individual-pops
branch
Added more robust handling of edge case whereby crowd distance calculations should have distances of 0 for individuals in a rank by themselves. (I.e., they have no neighbors, so no distance.)
This resolves an issue found by Robert Patton while he was using NSGA-II.