PostNL

Transcript

1. TRAP Tactical Route Assignment Planning Tim Jonker Data Scientist Fleur Doorman Data Scientist Daan Hiemstra Data Scientist Robin Hunteler Engineer Elise van Dam MLOps Mitch Vonk MLOps September 12 2023
2. 2 Tim Jonker Data Scientist @ PostNL
3. 3 Visit every address every day, no storage > 1 million parcels per day
6. Potential savings > 5 million/year By using TRAP 6
7. How does TRAP save money? 7 Initial assignment of areas to routes 1 2 3 4 5 6 7 8 9 10 11 12 Months Parcel volume • Anticipates on fluctuating parcel volume • Automates the planning process • Minimizes operational changes
8. 8 Algorithm 1 A geographic data processing step required for route planning Algorithm 3 An algorithm that sorts the delivery areas within a route in delivery order TRAP Algorithm 2 An optimization algorithm for the assignment of areas to routes
9. 9 TRAP’s goals and restrictions Algorithm 2 An optimization algorithm for the assignment of areas to routes Three ordered goals: 1. Minimize the number of routes 2. Maximize compactness of each route 3. Fair share of work for each route 5 ✓ Restrictions: • Complete assignment • Capacity • Consistency • Capabilities • Non-crossing
10. 10 Algorithm description Restrictions: • Each area should be assigned to a route • Maximum amount of work hours, parcels, stops, volume • Routes maximally deviate x% between iterations • Routes possess the required capabilities (Electric, Permission, Knowledge) • Routes can traverse another route for maximally y minutes Three ordered goals: 1. Minimize the number of routes 2. Maximize compactness of each route 3. Fair share of work for each route ✓ 5
11. 11 Available data Internal data: • Area characteristics • Route characteristics Open source data: • Travel times • Area shapes
12. 12 Areas are adjacent if there is a road between them that does not cross another area Transfer problem into two graphs TravelTimeGraph: Full connected graph with travel time as edge weights (OSRM) a b c d f e Adjacency graph: Full connected graph with travel time as edge weights (OSRM) Except for adjacent areas: 0 0 b 0 d 0 0
13. Generic term Consequence Graphical representation Capacitated restriction Maximum worktime, volume etc. Minimum Spanning Tree Promote adjacent areas Minimum ball Promote compact routes Spanning Forest Multiple routes 13 Decomposition of the problem definition Capacitated Minimum Ball Shaped Spanning Forest Combined
14. 14 Large Neighbourhood heuristic 1. Start with current assignment 2. Execute operations within a randomly selected neighbourhood 3. Terminate at stop criterion 4. Remove a route if solution is feasible or add a route if it is infeasible 5. Repeat 2-5 until no improvement
15. Large Neighbourhood heuristic 15 Iteratively search for neighbouring solutions that improve on the incumbent solution Neighbourhoods: • Insertion • Greedy • K-regret • Assignment • Swap • Edge removal • Shotgun removal
16. 16 Insert an unassigned area into a route Insertion Greedy Regret
17. 17 Assign one area to a different route Assign
18. 18 Swap two areas from different adjacent routes Swap
19. 19 Unassign areas at the edge of 2 adjacent routes Edge area removal
20. 20 Unassign areas at that are shot Shotgun removal
21. 21 Work in progress Results • TRAP requires fewer routes than currently used • TRAP results in more compact routes
22. 22 Algorithm 1 A geographic data processing step required for route planning Algorithm 3 An algorithm that sorts the delivery areas within a route in delivery order TRAP TRAP: Potential savings > 5 million/year Algorithm 2 An optimization algorithm for the assignment of areas to routes • Anticipates on fluctuating parcel volume • Automates the planning process • Minimizes operational changes
23. Questions? 23