-
Feature Request
-
Resolution: Done
-
Major
-
6.3.0.Final
-
None
-
NEW
-
NEW
Up to a few days ago I was convinced that nearby selection needs to cache all distance combinations.
But as that would make solving 100k VRP impossible on my machine (100k * 100k uses 38GB RAM), I've been looking at alternatives:
- Simply limited nearby till the n nearest isn't possible for CH's (as those 100 nearest might not be initialized).
- However, for CH's we can limit nearby to the n nearest initialized entities.
This splits up the nearby implementation, based on being used in LS or CH.
I 'll also allow a "shared" way, for small cases, where the entire nearby distance matrix can be cached.
- is duplicated by
-
PLANNER-362 Nearby Selection on VRP with 10k+ locations grabs a lot of memory for Local Search
- Resolved
- is related to
-
PLANNER-491 Partitioned Search
- Resolved
-
PLANNER-499 Use k-d tree for VRP: a more efficient datastructure to calculate nearby distances
- Open
- relates to
-
PLANNER-495 Multiple nearby selectors should share their distance matrix
- Resolved