An Integrated Approach to Load Matching, Routing, and Equipment Balancing Problem

1. An Integrated Approach to Load Matching, Routing, and Equipment Balancing Problem Sarah Root June 8, 2005 Joint work with advisor Amy M. Cohn

2. Overview • Problem description • Load matching, routing and equipment balancing problems • Literature review • Modeling approaches • Traditional multi-commodity flow model • Alternative cluster-based modeling approach • Implementation details • Initial computational results • Future research goals and directions

3. HFNO HFNO Planning Process • Load planning (Package routing) • Determine routing or path for each package • Service commitments and sort capacities must not be violated • Load matching (FeederOpt) • Match loads together to leverage cost efficiencies • Assign volume to a trailer type • Equipment balancing • Delivering loads from origin to destination causes some areas of the network to accumulate trailers and others to run out • Redistribute trailers so that no such imbalances occur • Driver scheduling (FSOS) • Take output of load matching and equipment balancing problems and assign drivers to each tractor movement • Load planning (Package routing) • Determine routing or path for each package • Service commitments and sort capacities must not be violated • Load matching (FeederOpt) • Match loads together to leverage cost efficiencies • Assign volume to a trailer type • Equipment balancing • Delivering loads from origin to destination causes some areas of the network to accumulate trailers and others to run out • Redistribute trailers so that no such imbalances occur • Driver scheduling (FSOS) • Take output of load matching and equipment balancing problems and assign drivers to each tractor movement

4. Current Solution Approach • Ensuring a feasible solution to the problem is difficult! • Minimizing cost is even harder!!! • Assume all data is deterministic • Break problem into sequential problems • Solve each problem individually • Renders problem tractable • Fails to capture interaction between different levels of the planning process, negatively impacting solution quality

5. Research Goals • Improve solution quality by integrating different steps of the planning process • Develop novel modeling approaches and algorithms to exploit underlying problem structure • We have initially integrated the load matching, load routing,and equipment balancing stages of the planning process (LMREBP) • Allows loads and empties to be routed together to realize cost savings • Improve solution quality by integrating different steps of the planning process • Develop novel modeling approaches and algorithms to exploit underlying problem structure • We have initially integrated the load matching, load routing,and equipment balancing stages of the planning process (LMREBP) • Allows loads and empties to be routed together to realize cost savings

6. Load matching problem i j i j i j Planning Process cijs • Assume all volume is assigned to 28’ trailers • Non-linear cost structure: single trailer combination vs. double trailer combination • Each load must be delivered to its destination within its time window cijs cijs cijd cijs ≤ cijd ≤ 2cijs

7. Planning Process • Load routing problem • Tractors can stop at intermediate nodes to reconfigure the trailers it pulls • Time feasibility—travel time and allowances • Relationship between load matching and load routing • Which loads should be matched together? How should these loads be routed? • Strongly interconnected—both decided simultaneously

8. Planning Process • Equipment balancing problem • Some nodes in the network have more inbound loads than outbound loads; these nodes accumulate trailers • Some nodes in network the have more outbound loads than inbound loads; these nodes run out of trailers • How should empty trailers be redistributed such that each node in the network is balanced? • Assume trailers do not have time windows

9. Literature Review • Specific LMREBP not considered in the literature to the best of our knowledge • Bodies of related literature • General multi-commodity flows • Multi-commodity flows with non-linear arc costs • Express package industry • Time windows • Empty balancing • See prelim proposal for a more detailed literature review

10. Traditional Approach to Modeling LMREBP • LMREBP is at its core a multi-commodity flow (MCF) problem • More difficult than a traditional MCF problem • Time windows • Non-linear cost structure • Network size • 2,500 nodes; 24,000 arcs; 15,000 commodities routed daily in the United States network • Traditional MCF formulation can be modified to capture the LMREBP

11. Traditional MCF Formulation • Let a node j represent a location andtime • Define the following variables: • xijk = number of commodity k flowing on arc (i,j) • yij = number of empty trailers flowing on arc (i,j) • sij = number of single loads flowing on arc (i,j) • dij = number of double loads flowing on arc (i,j)

12. Traditional MCF Formulation • Define the following parameters: • cijs =the cost of a single load flowing on arc (i,j) • cijd=the cost of a double load flowing on arc (i,j) • bjk =supply or demand of commodity k at node j • bjk> 0 if node j has a supply of commodity k • bjk < 0 if node j has a demand for commodity k • A=the set of all arcs (i,j) • V=the set of all nodes j • F=the set of all facilities f • K=the set of all commodities k • Vf =the set of nodes corresponding to facility f

13. Traditional MCF Formulation min cijs sij+ cijd dij s.t. xjik - xijk = bjk j in V, k in K sij + 2dij =xijk + yij  (i,j) in A  ( bjk +  yji -  yji ) = 0f inF xijk, yij, sij, dij in Z+ (i,j)єA (i,j)єA i:(j,i)єA i:(i,j)єA kєK jєVf kєK i:(j,i)єA i:(i,j)єA

14. Traditional MCF Formulation • Large number of constraints (|V||K|+|A|+|F|) • Huge number of nodes and large number of commodities! • VERY fractional LP relaxation • Incentive to send ½ double trailers instead of single trailers because of cost structure • Problem does not converge to a feasible solution even after relaxing time requirements • Motivates the need for an alternative modeling approach

15. Alternative Cluster-based Approach • Instead of considering the movement of trailers along each arc, consider groups of trailers which move together • A cluster is a set of loads, a set of empties, the routes they take, and the tractor configurations that pull them • Every load in the cluster moves completely from origin to destination • Only define clusters which are feasible

16. Alternative Cluster-based Approach

17. Alternative Cluster-based Approach

18. -2 2 2 -2 Alternative Cluster-based Approach

19. Alternative Cluster-based Approach • Let a node represent a facility (i.e., physical location) • Define the following variables: • xc = the number of cluster c used • Define the following parameters: • cc = cost of cluster c • cl = 1 if cluster c contains load l; 0 otherwise • cf = the impact of cluster c on trailer balance at facility f • L = the set of all loads l • F = the set of all facilities f

20. Alternative Cluster-based Approach min  cc xc s.t. cl xc = 1  l in L cf xc = 0  f in F xc in Z+ c c c

21. Alternative Cluster-based Approach • For a given cluster, it is easy to identify cost and whether or not it is time feasible • |L|+|F| constraints • Much stronger LP relaxation • Moves as part of a ½ double combination prohibited where both loads do not exist • Converges quickly to an integer solution • Flexibility in further expanding the problem scope • Non-facility meets, triple trailers, allowance times

22. Alternative Cluster-based Approach • Though the number of variables is very large, there are a number of ways to address this obstacle • Feasibility • Huge number of ways in which loads and empty trailers can be combined • Many of these ways are infeasible • Particularly true in complex clusters—time to reconfigure tractors and drive circuitous mileage can lead to violation of loads’ time windows • Dominance • Many ways to combine a given set of loads • Each potential combination corresponds to the same column in the constraint matrix • We only need to consider the cluster with the cheapest cost, as it would be chosen in an optimal solution

23. + = Alternative Cluster-based Approach • Indifference • Minimally dependent—cannot be broken into smaller pieces • For all sets of trailers T’T and T’’T such that T’T’’ =  and T’ T’’ = T, there is at least one leg containing both a trailer from T’ and T’’ if |T’|,|T’’| > 0

24. Implementation Details • Use of “cluster templates” to create initial clusters • Leverage the idea of dominance

25. Implementation Details

26. Implementation Details • In the data we’ve been given, nodes represent sorts • Multiple sorts can occur in the same building throughout a day; these will correspond to multiple nodes • We balance each building, not each node! • This node definition can limit matching opportunities when we use cluster templates

27. Travel time 253 Load 1880 Load 1882 Orig: 425 Dest: 411 ED: 2192 LA: 2476 Orig: 425 Dest: 412 ED: 2192 LA: 2800 Implementation Details • Consider an example Building 129 Building 93

28. 425 411 412 1882 1880 1880 1882 1882 93 129 Implementation Details • We need to fit these loads into a cluster template • Using original nodes • Using building numbers

29. Implementation Details • Three legs in the cluster if we consider moves between node numbers • Single leg in the cluster if we consider moves between building numbers

30. Implementation Details • Benefits in initially considering moves between building numbers • More opportunities to match loads together in clusters • ~10,000 clusters using original node numbers • ~50,000 clusters using building numbers • Equipment balancing is done at the building level, not the node level

31. Computational Results • Moderately sized data set • 2,000 loads; 600 nodes; 250 buildings • Time windows extremely tight for the loads • Lower bound on the optimal objective—$482,849 • Route each load directly from origin to destination at cost of half a double combination • Balance the empties in the system using a transportation problem where arc costs are for half double combinations • Add cost of routing loads and balancing empties

32. Computational Results • Traditional MCF model did not converge to an integer feasible solution • Alternative formulation converged to an integer solution within 15 seconds • Within 2 minutes, within 1% of the optimal solution using the subset of clusters • Still incentive to split empties into ½ double empty combinations—does not converge to a provably optimal solution in 2 hours • UPS solution without load routing—$609,854 • UPS solution with load routing—$585,128 • Michigan solution—$584,881

33. Computational Results • More than 75% of trailers move at least one leg as part of a double combination • Solution will improve as we add new cluster templates • Addition of a single cluster template has saved $5,000 in this network • Ideas for cluster templates are appreciated! • Extremely tight time windows in the data set we’ve been using • Looser time windows in the US network allow for more potential to match loads and realize cost savings

34. Future Research Directions • Finish up this portion of the research • Include new cluster types when solving the problem • Leverage “symmetry” of the problem • Investigate sensitivity of solution to time windows • Move to larger networks (e.g., US network) • Use dual information to generate promising clusters—column generation • Further integrate steps in the planning process • Assigning volume to trailers

35. Summary • Integrated multiple steps in a complex planning process • Load matching, routing and equipment balancing • Developed a novel modeling approach that overcomes the traditional difficulties associated with traditional modeling approach • Framework can capture complexities associated with the real-world decisions to be made and allows us to extend the problem scope • Initial computational results are promising • Can quickly find a good solution • More matching opportunities in the US network

36. Questions?