Models for Minimizing Backhaul Costs through Freight Collaboration

Erin Bailey
Department of Civil and Environmental Engineering
West Virginia University
Room 621 ESB, P.O. Box 6103
Morgantown, WV 26506
E-mail: ebailey2@gmail.com

Avinash Unnikrishnan
Department of Civil and Environmental Engineering
West Virginia University
Room 621 ESB, P.O. Box 6103
Morgantown, WV 26506
E-mail: Avinash.Unnikrishnan@mail.wvu.edu
(corresponding author)

Dung-Ying Lin
Assistant Professor
Department of Transportation and Communication Management Science
National Cheng Kung University, 1 University Road, Tainan City, 70101,
Taiwan
E-mail: dylin@mail.ncku.edu.tw
Phone: +886 6 2757575 Ext. 53226
Fax: +886 6 2753882

Submitted for presentation at Transportation Research Board Annual Meeting and publication in
Transportation Research Record

Word Count:
Text: 6266 words (abstract:149)
Tables: 5 x 250 = 1250
Figures: 1 x 250 = 250
Total: 7766

Submission date: November 14th 2010

TRB 2011 Annual Meeting Paper revised from original submittal.

 ABSTRACT

The competitive nature in the trucking industry has forced trucking firms to develop innovative
solutions to improve their operational efficiency and decrease marginal costs. Recently, one way
carriers and shippers are accomplishing this is through collaborating their operations in various
collaboration strategies. This work focuses on how a carrier can introduce collaboration in their
operations by fulfilling a collaborative carrier’s or shipper’s delivery requests on its backhaul
route. Two optimization models are developed to route the carrier of interest’s backhaul routes
and select collaborative shipments to fulfill; one is formulated as an integer program and the
other is formulated as a mixed integer program. Two solution methodologies, a greedy heuristic
and tabu search, are used to solve the two problems, and numerical analysis is performed with a
real world freight network. Numerical analysis reveals that the percentage of cost savings for
backhaul routes can be as high as 27%.

TRB 2011 Annual Meeting Paper revised from original submittal.

 Bailey, Unnikrishnan, and Lin 1

1 INTRODUCTION
2 The U.S. economic strength and growth depends significantly on the movement of goods and
3 particularly the trucking industry for its ground movement. The trucking industry has become
4 more competitive and dynamic since the deregulation of the freight industry in the 1980s. Also,
5 continued rises in highway congestion have caused the government to encourage more freight to
6 shift from being shipped by truck to rail (1). Consequently, it is vital for trucking firms to
7 improve their operations and lower their marginal costs through innovative solutions. It is
8 essential for small and medium-sized firms to decrease their operation costs in order to compete
9 with the larger ones. Collaboration among trucking firms is a great way for the trucking industry
10 to help accomplish these goals. Freight demand is often geographically widespread, creating
11 empty backhauls after shipment delivery, called deadheading. Minimizing deadheading can
12 reduce operational costs and increase efficiency. When carriers form alliances they can
13 collaborate their truck operations and change non-profitable empty hauls into profitable
14 backhauls. Reducing deadheading also lessens the environmental impact of truck freight
15 transportation, by way of less fuel consumption and lower emissions. Collaboration can also
16 help them increase asset utilization, decrease lead times, and improve their service levels.
17 In the last 20 years, collaboration has occurred between the various freight companies
18 involved in the (full) truckload industry. Two common types of collaboration in truckload
19 transportation are shipper collaboration and carrier collaboration. Shipper collaboration involves
20 multiple shippers coordinating with one or more carriers and matching their loads so that optimal
21 routing can be found to execute all of the shipments among the shippers at minimum cost. In
22 truckload carrier collaboration, multiple carriers focus on reducing the operating costs by pooling
23 delivery tasks and vehicle capacities in common to find efficient ways of fulfilling all shipping
24 requests.
25 Collaboration among agents involved in the less-than-truckload (LTL) industry has only
26 been recently studied and primarily focuses on the collaboration among carriers in the LTL
27 industry. LTL transportation involves shipping smaller freight, typically from many shippers
28 and has shorter planning periods than truckload transportation. Because many LTL carriers have
29 overlapping networks and use distribution centers, depots or warehouses, it provides the
30 opportunity for carriers to easily exchange freight and utilize one another’s capacities (2).
31 In this paper we study how small- to medium-sized truckload carriers can reduce empty
32 backhauls to a depot or distribution center location by adding pickup and delivery tasks of
33 collaborative shippers or carriers to their backhaul. A carrier of interest can increase their profit
34 margins by fulfilling a shipping request of another firm during its backhaul route. Reducing the
35 empty and often long hauls to the depot will greatly increase the carrier’s profitability. We
36 develop two mathematical models in this paper to route backhaul shipments for the carrier of
37 interest’s fleet of trucks. The first model allows at most one collaborative shipping request be
38 fulfilled for each truck and is formulated as an integer program. Model 2 is more complicated
39 and allows multiple pick-up and delivery tasks for each truck and is formulated as a mixed
40 integer program. Model 1 is solved using a greedy heuristic and a Tabu search heuristic is
41 provided for solving Model 2. If it is feasible for the carrier of interest to fulfill another shipper’s
42 request during its backhaul route, the routing is as follows: from a truck’s last delivery location,
43 the empty truck would pick up a collaborative shipper's or carrier’s shipment, deliver the
44 shipment to its destination, then return back to the depot. One advantage of our model approach
45 over a full collaboration is that the carrier of interest would not have to alter its existing linehaul
46 routes, and only slightly change its backhaul routes. This type of collaboration can be a good

TRB 2011 Annual Meeting Paper revised from original submittal.

 Bailey, Unnikrishnan, and Lin 2

1 transition into full collaboration later, where all delivery tasks and vehicle capacities are in
2 common and optimal routing is found for the collaborative network. The modeling of the latter
3 type of collaboration is outside the focus of this paper. Note that the models we present are
4 flexible with respect to whom the carrier of interest collaborates. Collaboration can occur with
5 other TL shippers that have origin-destination shipping locations close to the carrier of interest’s
6 backhaul routes. Collaboration with carriers is also possible if a carrier does not have available
7 capacity to service all of their requests. The models described in further sections refer to the
8 collaborative shipper or carrier as the “collaborator” or “collaborative freight company” to not
9 exclude a type of collaboration.
10 The remainder of the paper is organized as follows. A detailed literature review is
11 provided in the next section. The two formulations are then provided. The two solution
12 methodologies are then described. Numerical analysis is conducted on a real world network with
13 modified parameters, followed by the conclusions and directions for future research.
14
15 LITERATURE REVIEW
16 One group of carrier collaborative literature studies truckload transportation procurement, where
17 least cost methods of obtaining additional capacity for a carrier of interest or shipper of interest
18 are investigated (3,4,5,6). Although not distinctively collaboration, Chu (3) and Ball et. al. (4)
19 studied how a shipper with limited capacity could execute all of its delivery tasks by utilizing an
20 outside carrier to meet the demand it could not self fulfill. Song et. al. (5) and Figliozzi (6)
21 studied an auction-based carrier collaboration mechanisms among TL carriers. Hernández et. al.
22 (2) studied LTL carrier collaboration and developed the single carrier collaboration problem
23 (SCCP) to gain insights on the potential for a carrier with excess demand to collaborate with
24 other carriers to service that demand. Their results revealed savings up to 59% under a
25 collaborative scenario when compared to a leasing option to obtain additional capacity.
26 Another set of literature addresses LTL collaboration from the perspective of optimizing
27 the operations of the whole collaborative freight network (7,8). Dai and Chen (7) developed a
28 model for LTL collaboration that could be applied to a set of shippers and/or carriers. They
29 examined this in a scenario where all carriers’ delivery tasks were shared and all vehicle
30 capacities were known. Their model allowed for pick-up and/or delivery at any node in the
31 network. Berger and Bierwirth (8) developed a model to optimize the operations of all carriers
32 while ensuring that no carrier would lose profit under the collaboration. They used a cash flow
33 model, called the Collaborative Carrier Routing Problem, to reassign requests to a carrier in
34 which the profit of the collaborative network was maximized.
35 Shipper collaboration is limited in literature and was first introduced by Ergun, Kuyzu
36 and Savelsburg (9). They studied how truckload shippers can collaborate to minimize asset
37 repositioning, thereby reducing deadhead trips. They translated their problem into the Lane
38 Covering Problem (LCP), where a minimum cost set of constrained cycles is found to cover a
39 subset of arcs (delivery lanes) in a directed Euclidean graph. They presented algorithms to solve
40 this problem, which were found to be polynomially solvable. Özener and Ergun (10) developed
41 cost-allocation schemes in similar shipper alliances after determining the optimal collaborative
42 routing. Using the basis of the LCP, Ergun et. al. (11) also developed optimization technology to
43 assist in identifying collaborative shippers’ repeatable, dedicated truckload continuous move
44 tours in order to reduce the need for carriers to reposition empty trucks. Ergun et. al. (12, 13, 14)
45 have addressed carrier collaboration in the air cargo and sea cargo industries and have developed
46 models and allocation mechanisms that are applicable to these types of collaborations.

TRB 2011 Annual Meeting Paper revised from original submittal.

 Bailey, Unnikrishnan, and Lin 3

1 Other literature focuses on minimizing empty truckloads for a single company’s

2 operations without collaboration. Arcelus et. al. (15) formulated a dynamic model for reducing
3 empty hauls within a company’s hub-and-spoke structured network. They performed sensitivity
4 analysis on whether or not it was beneficial to obtain more requests for specific origin-
5 destination pairs. Jordan (16) examined deadheading trip reduction for a single company’s
6 operations with more than two terminals and developed a Lagrangian Relaxation method for
7 determining the optimal backhaul routing strategy. The research presented in this paper
8 differentiates itself from past works as it focuses on optimizing the backhaul route of a specific
9 carrier of interest through collaboration with other shippers and/or carriers. We also ensure that
10 the new modified backhauls are not significantly longer than the existing backhauls so that the
11 carrier of interests can satisfy the new requests without significantly modifying their driver
12 schedule. Carriers can use the models presented in this paper to evaluate the savings obtained by
13 optimizing the backhauls in their entire network or part of their network. The integer
14 programming and mixed integer programming formulations of the two models proposed in this
15 work are described next.
16
17 MODEL FORMULATIONS
18 The problem of interest is to reroute the empty backhaul trucks of a carrier to fulfill other
19 collaborative carriers’ or shippers’ requests. Given a set of locations for the trucks’ last
20 deliveries (described heretofore as origin locations), the models decide if a collaborative freight
21 operator’s request can be fulfilled by one of the trucks. If not, the truck will keep its original
22 empty backhaul route. The two models presented in this section route the backhauls such that
23 the cost to the carrier of interest is minimized. Both models consider truck capacity constraints
24 and vehicle driver hour constraints. The backhaul trip length cannot exceed an upper time limit,
25 which is explained in more detail in the numerical analysis section. The first model, called
26 Model 1, is an integer programming model that pairs each origin node with a collaborator node
27 or the depot node. If an origin is paired with a collaborator, then it will fulfill that shipping
28 request before returning to the depot. This model will only allow at most one collaborative
29 shipping request be fulfilled for each truck. Model 1 is similar to a constrained matching
30 problem. The second model, Model 2, is a mixed integer programming problem and a more
31 complicated model than Model 1. Rather than a matching problem, Model 2 is similar to a
32 capacitated vehicle routing problem with backhauls (17, 18) with the vehicles not returning to
33 the originating depot. Note that in the above mentioned works both the line haul and backhaul
34 deliveries are optimized whereas in this work we focus on the backhaul variation. This model
35 also has the advantage of allowing multiple pick-up and deliveries in one truck’s backhaul.
36 Similar problems involving pick-up and deliveries have been studied in the literature by Dumas
37 et al and Nagi et al (19, 20). Model 2 is different as there are constraints on the route length
38 corresponding to driver hour restrictions and there is the option of not picking up goods if it is
39 not profitable for the carrier of interest. The model also ensures that an O-D shipment is not split
40 up in its delivery.
41
42
43 Model 1
44 Model1 maximizes the collaborative savings, sij, of the carrier of interest. Some explanation of
45 the collaborative savings computation is needed before the model formulation can be presented.
46 The original transportation cost, cid, is the cost for the truck to go from an origin node i to the

TRB 2011 Annual Meeting Paper revised from original submittal.

 Bailey, Unnikrishnan, and Lin 4

1 depot node d and cip, is the cost for the truck to leave its origin node, carry out the shipping
2 request of collaborator p and return to the depot. If a collaborator’s request is fulfilled, then that
3 collaborator will provide compensation to the carrier of interest, resulting in a revenue rp for that
4 request. The backhaul savings, as shown in equation (1), is computed by subtracting the
5 collaborative transportation cost from the original transportation cost. Equation (2) explains that
6 the backhaul savings is zero for a truck returning to the depot instead of making a backhaul
7 delivery.
8 sip = cid –(cip – rp) (1)
9 sid =cid – cid = 0 (2)
10
11 The model parameters and the variables used in the Model 1 formulation are described below.

12 Model 1 Parameters:
13 O set of all origin nodes i
14 P set of all collaborators p
15 J set of all collaborator nodes p and the depot node d, indexed by j
16 tij total trip length, in time, to travel from the origin node, pick up and deliver request p, and
17 return to depot
18 Ti upper limit of the total backhaul trip length, in time, for origin node i
19 uj demand of node j
20 C capacity of trucks
21
22 Model 1 Variable:

23
24 The formulation is given as follows:
aximize

25 Subject to:

26
27
28 The objective function of Model 1 maximizes the collaborative savings of the carrier of
29 interest. Constraint (4) ensures that each origin truck node is paired with exactly one of the
30 collaborators p or the depot d. Contraint (5) ensures that each collaborator is paired with no
31 more than one origin node. Constraint (6) ensures that the total backhaul trip length does not
32 exceed the maximum trip length for a truck at node i so that the carrier of interest can meet the

TRB 2011 Annual Meeting Paper revised from original submittal.

 Bailey, Unnikrishnan, and Lin 5

1 collaborative freight operators shipping request without significantly altering the driver schedule.
2 Note that the demand at the depot node is zero and the capacity of all trucks is assumed to be
3 equal. Therefore there is no need for an explicit capacity constraint as the list of collaborator
4 nodes can be pre-processed to remove all collaborators whose demand is greater than the truck
5 capacity.
6
7 Model 2
8 The parameters and decision variables for Model 2 are described, followed by the mixed integer
9 programming formulation.
10
11 Model 2 Parameters:
12 N set of all nodes i
13 A set of all arcs (i,j)
14 P set of all collaborators’ pick-up nodes
15 B set of all collaborators’ delivery nodes
16 O set of all origin nodes for empty trucks
17 D depot node
18 M set of all collaborator’s origin-destination shipment pairs m
19 cij transportation cost to travel on arc (i,j)
20 ri carrier of interest’s revenue for fulfilling a collaborator’s shipment from pick-up node i
21 uim collaborator’s product demand of O-D shipment m at node i
22 tij length of time to travel on arc (i,j)
23 To upper limit on length of backhaul trip, in time, of truck originating from node o
24 C capacity of trucks
25
26 Model 2 Variables:

27
28 yijom quantity of product m shipped on arc (i,j) by truck originating from node o
29

30
31 The formulation is given as follows:
32
inimize

33
34
35 Subject to:

36

TRB 2011 Annual Meeting Paper revised from original submittal.

 Bailey, Unnikrishnan, and Lin 6

4
o

10

11

12
13 The objective function for Model 2 subtracts the total revenue earned from collaborative
14 shipments from the total transportation cost. Therefore, the objective function minimizes the
15 overall cost of the carrier of interest’s backhaul trips. Constraints (9) and (10) are truck flow
16 conservation constraints for the collaborators’ pick-up and delivery nodes. Constraint (10)
17 ensures that at most one truck visits the pick-up nodes, thereby ensuring that an O-D shipment is
18 not split. Constraint (9) ensures that the number of trucks leaving and arriving at a collaborator’s
19 node is equal. Therefore, constraints (9) and (10) ensure that the number of trucks arriving and
20 leaving a collaborator’s node is either zero or one. Constraints (11) and (12) are truck flow
21 constraints for the empty trucks at the origin nodes, ensuring that exactly one trip leaves each

TRB 2011 Annual Meeting Paper revised from original submittal.

 Bailey, Unnikrishnan, and Lin 7

1 origin node and no trips arrive at the origin nodes. Constraints (13) and (14) are truck flow
2 constraints for the depot node, ensuring that each trip returns to the depot node and no trips leave
3 it. Constraint (15) ensures that the product demand at all of the nodes is met; the demand at the
4 collaborator nodes is only met if the nodes are visited. The demand for the origin and depot
5 nodes is zero. It also ensures that each collaborator’s shipment is delivered to its corresponding
6 delivery node because the demand parameter has an O-D shipment type specification. Constraint
7 (16) ensures that a backhaul route’s total trip length does not exceed the maximum allowable trip
8 length. This places a restriction on trucks to not fulfill those collaborator’s shipments that will
9 extend the backhaul trip beyond the maximum allowable driver hours. Constraint (17) ensures
10 that the capacity of a truck is not exceeded; all trucks have equal capacity in this model.
11 Constraint (18) must be included to eliminate any subtours in the solution and is the well-known
12 generalized subtour elimination constraint (21). The next section describes the heuristics to solve
13 the above formulations.
14
15 SOLUTION METHODS
16 We present a greedy heuristic to solve Model 1 and a Tabu search based heuristic to solve Model
17 2. Model 1 and 2 have integer variables and are difficult to solve for larger sized problems. Also
18 Model 2 has the subtour elimination constraints which explode for any decent sized network.
19 The greedy algorithm to solve Model 1 is described in the next subsection followed by the Tabu
20 search heuristic.
21
22 Greedy Algorithm
23 Many weighted matching problems can be solved using a “greedy” heuristic methodology to
24 obtain good solutions in a reasonable amount of time (16). Therefore, we use a greedy matching
25 algorithm to solve Model 1. Let A = {(iO, jJ)} denote the set of arcs, where O is the set of
26 all origin nodes and J is the set of all collaborator nodes and the depot node. The algorithm is as
27 follows:
28 (1) Set all xij = 0 for all arcs (i,j).
29 (2) Delete any (i,j) from list A where the demand at j exceeds the vehicle capacity, C.
30 (3) Select the arc (i,j) from list A which has the maximum value of arc savings, sij. (sij must
31 be greater than or equal to zero). Remove the selected (i,j) from list A.
32 (4) For the selected (i,j) in step (3), if xij tij  Ti (the total trip length does not exceed the
33 maximum trip length from origin i), then set xij = 1, remove any arc (i,j) from list A that
34 contains either the selected i or selected j, except where the selected j is the depot node D.
35 If xij tij > Ti, go back to step (3).
36 (5) Repeat steps (3) and (4) until number of matched pairs (xij = 1) equals the number of
37 origins, O.
38 The greedy algorithm might not produce optimal solutions to Model 1, but it provides good
39 solutions for examining numerical results of the model. The Tabu search methodology is
40 described next.
41
42
43 Tabu Search Methodology
44 A brief overview of the Tabu search methodology used for solving Model 2 is provided here.
45 Tabu search was chosen among other meta-heuristics as it was found to be extremely efficient
46 and thus popular for solving vehicle routing problems (22,23,24). While the focus of the current

TRB 2011 Annual Meeting Paper revised from original submittal.

 Bailey, Unnikrishnan, and Lin 8

1 paper is on a new problem definition and formulation, the design of the Tabu search heuristics
2 follows the common principles (25).

1. Initialization

2. Create a candidate list of

No
moves

4. Is aspiration criteria
satisfied?
Yes 3. Is a candidate tabu?

No

Yes 5. Evaluate the candidate

6. Is the candidate better than

the incumbent?
No

Yes

7. Update the incumbent and

No
place the candidate into tabu list

8. Move to the candidate

9. Is the stopping criterion

met?

Yes

10. Terminate and report the

incumbent solution
3
4 FIGURE 1: Tabu search flowchart
5

TRB 2011 Annual Meeting Paper revised from original submittal.

 Bailey, Unnikrishnan, and Lin 9

1 Step 1 initializes the Tabu search procedure by assuming that no collaborator’s request is
2 fulfilled. Further, initial parameters (the size of Tabu list, maximum number of iteration,
3 memory size and incumbent solution) are also assigned in that step. Step 2 creates a candidate
4 list of moves where each move corresponds to a new solution generated from the current
5 solution. Details of the neighborhood search and the edge-swapping procedure for generating
6 candidate lists are available in Glover (25). Step 3 selects a candidate move from the candidate
7 list and checks if the selected candidate is tabu. If a candidate is tabu, go to Step 4. Otherwise,
8 the procedure evaluates the performance of the move based on the objective function (Equation
9 (8)) in Step 5. In Step 4, if a candidate satisfies aspiration criteria (yields better objective value
10 than the incumbent solution), the procedure goes to Step 5. Otherwise, the procedure goes back
11 to Step 2 and Step 3 to choose another candidate move. If the candidate move evaluated in Step
12 5 yields better objective value, the procedure goes to Step 7 to update the incumbent solution and
13 places the move into the tabu list such that the move will not be admissible in the following few
14 iterations. The procedure then moves to the candidate in Step 8 and checks the stopping criteria
15 in Step 9. Two stopping criteria are employed in this research. The first one is the iteration limit
16 which constrains the maximum number of iterations that can be performed. However, if the
17 incumbent solution fails to improve over the pre-defined number of iterations, the procedure is
18 terminated, which forms the second stopping criterion. If any of the stopping criteria is met, the
19 procedure stops and the incumbent solution is reported. Otherwise, the procedure goes back to
20 Step 2 and starts the procedure over again.
21
22 NUMERICAL ANALYSIS
23 Computational runs are conducted to analyze the variation in model savings obtained through
24 collaboration using Model 1 and Model 2. The example network used in this study is first
25 described followed by the results from the computational runs.
26
27 Example Network Setup
28 An example network was used to implement Model 1 and Model 2 and to analyze the models’
29 performance and potential cost savings for the carrier of interest. Real network data from a
30 third-party logistics firm was used to establish the carrier of interest’s inputs into the model.
31 This data set includes 15 backhaul origin nodes and one depot node location. The depot node
32 location is York, PA and the origin node locations are the following: Orlando, FL; Crofton, MD
33 locations; Clementon, NJ; Glassboro, NJ; Robinsville, NJ; Olean, NY; Oneonta, NY;
34 Miamisburg, OH; Troy, OH; Allentown, PA; Bensalem, PA; Hazleton, PA; Palmyra, PA; and
35 Norfolk, VA. The maximum backhaul trip length, Ti in Model 1 and To in Model 2, was
36 determined by considering the Federal Motor Carrier Safety Administration hours-of-service
37 regulations for truck drivers (26). When fulfilling collaborator’s shipping requests, the carrier of
38 interest’s backhaul trip should not be lengthened such that any driver’s hours would exceed
39 limits set by FMCSA regulations. The time constraints used for this model in hours for the
40 origins are: 22.5, 13, 6.75,6.75, 5.5, 9.75, 4.75, 9.25, 20, 6,7, 8.75, 6, 7.75, 10.25 respectively.
41 To obtain potential collaborator pick-up and delivery locations and their O-D demand, we used
42 shipping lanes from a real shipping network. We chose 45 of the O-D pairs as potential
43 collaborators due to their geographic proximity to the carrier of interest’s backhaul route
44 network. Table 1 provides the list of potential collaborator nodes with the demand units. Due to
45 privacy constraints, the trucking company did not reveal the nature of the goods and modified the

TRB 2011 Annual Meeting Paper revised from original submittal.

 Bailey, Unnikrishnan, and Lin 10

1 demands. The truck fleet is considered homogenous for our example with a capacities of 1000
2 demand units.
3
4 TABLE 1 Collaborator Nodes
Product Product
Pick-up City Delivery City Demand Pick-up City Delivery City Demand
Units Units
Jacksonville, FL Charlotte, NC 672 Columbus, OH Harrisburg, PA 360
Jacksonville, FL Rocky Mount, NC 328 Sidney, OH Harrisburg, PA 1083
Haines City, FL Rocky Mount, NC 1142 Sidney, OH Eighty Four, PA 745
Lake Wales, FL Charlotte, NC 323 Columbus, OH Harrisburg, PA 341
Lake Wales, FL Baltimore, MD 292 Springfield, OH Harrisburg, PA 901
Savannah, GA Charlotte, NC 157 York, PA Harrisburg, PA 429
Owings Mills, MD Harrisburg, PA 443 Quakertown, PA Harrisburg, PA 155
Baltimore, MD Harrisburg, PA 375 Coatesville, PA Harrisburg, PA 243
Gaithersburg, MD Harrisburg, PA 1270 Hermitage, PA York, PA 234
Hampstead, MD York, PA 221 Hershey, PA Harrisburg, PA 374
Mount Olive, NC York, PA 235 York, PA Harrisburg, PA 126
Cherry Hill, NJ Philadelphia, PA 388 York, PA Manayunk, PA 758
Cranbury, NJ Philadelphia, PA 400 Leola, PA York, PA 155
Trenton, NJ Philadelphia, PA 303 Manheim, PA Philadelphia, PA 4290
South Brunswick, NJ Pennsauken, NJ 576 Bethlehem, PA Manayunk, PA 225
Edison, NJ Philadelphia, PA 224 Breinigsville, PA Philadelphia, PA 299
North Brunswick, NJ Pennsauken, NJ 672 Hazleton, PA Philadelphia, PA 262
North Brunswick, NJ Philadelphia, PA 229 Plains, PA Harrisburg, PA 591
Totowa, NY Harrisburg, PA 192 Philadelphia, PA Harrisburg, PA 378
Canandaigua, NY Conklin, NY 1656 King of Prussia, PA York, PA 315
Canandaigua, NY Harrisburg, PA 823 Leesport, PA York, PA 342
Silver Springs, NY Kirkwood, NY 199 Richmond, VA Harrisburg, PA 239
Middletown, NY York, PA 667
5
6 The data set structure for the transportation costs, revenue, travel time, and product demand
7 varies from Model 1 to Model 2. For the Model 2 parameter, tij, the travel time between every
8 node in the network was determined and 1 hour for loading time and 1 hour for unloading time
9 was added to arcs containing pick-up or delivery nodes, respectively. The Model 2 data set for
10 transportation costs required finding the distance between each node in the network, and
11 applying a unit cost per mile. The base cost per mile used for both models’ networks was $1.60
12 per mile and is based on typical industry values. Additional data sets were established using
13 variations to the cost and the results are shown in the next section. The revenue parameter, ri, in
14 odel 2 represents the collaborator’s compensation for satisfying a shipment request from node
15 i. Each ri was taken as a percentage of the transportation cost, cij, from its pick-up node to its
16 delivery node. This revenue percentage was varied from 30% to 50% during the numerical
17 analysis of the models. The product demand input for Model 2 consisted of identifying product
18 demand for each O-D pair m, in which there was zero demand at every origin node and the depot
19 node, positive demand at each pick-up node and negative demand at each delivery node. Most of
20 odel 1’s data set can be taken from odel 2’s data set. The transportation time, tij, is not

TRB 2011 Annual Meeting Paper revised from original submittal.

 Bailey, Unnikrishnan, and Lin 11

1 needed between every node as in Model 2, but only from each origin node to each pick-up node,
2 each pick-up node to its corresponding delivery node, and from each delivery node to the depot
3 node. The loading and unloading time was added to the transportation time in Model 1 as
4 discussed earlier for Model 2. The collaborative savings, sij, data input of Model 1 is computed
5 from transportation costs and collaborative revenues as previously described in equations (1) and
6 (2). The transportation costs for the backhaul routes and revenue of each collaborator’s shipment
7 were found in the same manner as they were for Model 2. The input for product demand of
8 collaborator’s shipments is the same as in odel 2 but simplified in that there’s only one positive
9 demand value for each collaborator.
10
11 Model 1 - Analysis of Results
12 Table 2 summarizes the percent cost savings obtained from Model 1 and 2 with variations to the
13 transportation arc cost and the revenue percentage parameter. The fourth column, the total cost
14 of backhauls without collaboration, is the cost of all backhaul routes to travel from the origin
15 node to the depot node. The fifth column shows the total cost savings with collaboration and for
16 Model 1 this is the objective value solution. Computation times of running the models are also
17 provided in Table 2. The $1.60/mile transportation cost is considered our “typical” cost and
18 additional results are shown for transportation costs ranging from $1.20/mile to $2.00/mile. The
19 results of Model 1 show that varying the transportation cost does not affect the number of
20 matched backhauls to collaborators. The percent cost savings also does not vary with
21 transportation costs. For example, a $1.20/mile transportation cost at the 30% revenue level
22 results in 10 matched pairs and a collaborative cost savings of 13.17%; a $2.00/mile
23 transportation cost at the 30% revenue level produces these results. The savings dollar amount
24 increases with transportation cost increase because the transportation cost is built into the
25 computation of revenue and savings.
26 Variations to the percent revenue show that more backhaul routes are paired with
27 collaborators at a higher percent revenue than at a lower percent revenue. As shown in Table 2,
28 when the percent revenue is set at 30% and 40% of the transportation cost, 10 out of the 15
29 carrier of interest’s backhaul routes are matched with a collaborator’s shipment. When the
30 percent revenue is 50%, an additional 3 backhaul routes are matched with collaborator’s
31 shipments. The cost savings experienced for the backhaul routes are found to range from about
32 13% to 27% for the three revenue levels. More of the longer mileage backhaul routes match up
33 with collaborators at the 30% revenue level than the shorter mileage routes. Results show that
34 backhaul routes match up with collaborator O-D pairs that are very close geographically to the
35 backhaul route, which is as expected. Collaborator O-D shipments that consist of longer hauls
36 are more likely to be matched with a backhaul route than those with shorter hauls. This is
37 because the revenue, ri, that the carrier of interest receives is greater for the longer O-D hauls. It
38 should be noted that the assumption is made that all collaborators are willing to pay the
39 compensation price to the carrier of interest, be it 30%, 40% or 50% of the carrier of interest’s
40 shipping costs. Typically this price would be appealing to a collaborative shipper, but not
41 appealing for all collaborative carriers. Further research could explore the development of a
42 profit sharing mechanism incorporated with the models presented in this paper. The greedy
43 heuristic for Model 1 takes on an average 20% of the time needed to solve the tabu search
44 heuristic for Model 2 on a Dell Latitude Laptop with Core 2 Duo 2.53 GHz processor with 3 GB
45 Ram.

TRB 2011 Annual Meeting Paper revised from original submittal.

 Bailey, Unnikrishnan, and Lin 12

1 Affects of Model 1 results to variations of loading and unloading times at the collaborator
2 pick-up and delivery locations are shown in Table 4. The loading/unloading times for the base
3 data sets utilizes times of 1 hour for loading and 1 hour for unloading, which is typical of the
4 carrier of interest’s existing schedule. The results in Table 4 are from model runs all using the
5 30% revenue level and the base $1.60/mile transportation cost. For the example network,
6 reducing the loading/unloading times to 0.5 hour and 0.75 hour did not reduce the number of
7 matched pairs as compared to the 1 hour times. Extending the loading/unloading times to 1.25
8 hour and 1.5 hour reduced the number of matched pairs to 7 out of 15. The percent cost savings
9 is higher for the 1.25 hour case than the 1.5 hour case because the 7 matched pairs are not the
10 same 7 collaborators.
11 We also study the sensitivity of Model 1 results to reduction in the maximum backhaul time
12 length parameter. Recall that the maximum backhaul time length, Ti, is based on the FMCSA
13 hours-of-service rules (26). Results in Table 5 show the affects of reducing the original Ti values
14 by 10%. At the 50% revenue level, the reduced Ti results in only 9 of the 15 backhaul routes
15 matching with a collaborative shipment while the original Ti values have 13 of the 15 routes
16 matched.
17
18 Model 2- Analysis of Results
19 Model 2 was solved using the tabu search heuristic. After sensitivity analysis, the parameters
20 used in the tabu search were: (i) max number of iterations - 2000, (ii) size of tabu list - 7. Also
21 the program is terminated if the incumbent solutions fail to improve over 100 iterations. The
22 results of Model 2 are summarized in Tables 2 through 5. Table 2 shows the results of Model 2
23 when the transportation arc costs and the percent revenue parameters are varied. Model 2 results
24 show some similar trends to the Model 1 results. Model 2 results show that changing the
25 transportation arc cost while keeping the percent revenue constant does not affect the number of
26 backhaul routes matched to collaborators or affect the percent cost savings. The increasing of
27 the percent revenue parameter does not have as much of an effect on the number of matched
28 pairs as it does for Model 1. For the computational runs using $1.20/mile transportation cost,
29 only one additional backhaul route is matched when increasing the percent revenue from 30% to
30 50%. For the other transportation cost trials, the number of matched backhaul routes does not
31 increase with an increased percent revenue. Model 2 has more matched backhaul routes than
32 Model 1 at the 30% and 40% revenue levels, but Model 1 has more matched backhauls than
33 Model 2 at the 50% revenue level. However, all trials of Model 2 experience more percent cost
34 savings than Model 1. This is because some of the matched backhaul routes are matched to more
35 than 1 collaborator. Model 1 does not allow for a backhaul route to fulfill more than 1
36 collaborator’s demand. Table 3 shows how many collaborators each backhaul route is matched
37 with for varying transportation costs and revenues. Some of the backhauls are matched with two
38 collaborators, and the backhaul route from origin 9 is matched with up to 3 collaborators.
39 Table 4 summarizes how Model 2 is affected by variations to the loading and unloading time
40 parameters. The results in Table 4 are from model runs all using the 30% revenue level and the
41 base $1.60/mile transportation cost. Similar to Model 1 results, Model 2 results show that
42 increasing the loading and unloading times above 1 hour causes the number of matched
43 backhauls to decrease by 3 backhauls.
44 Model 2 results for reduction in the maximum backhaul length parameter, To, are
45 summarized in Table 5 and reveal similarities with Model 1 results. When the maximum

TRB 2011 Annual Meeting Paper revised from original submittal.

 Bailey, Unnikrishnan, and Lin 13

1 backhaul time length is reduced by 10% from the original times, the number of matched
2 backhaul routes decreases between 2 to 3 backhauls, depending on percent revenue.
3
4
5 TABLE 2 Results with Variations in Transportation Cost and Percent Revenue Parameters

Number of
Total cost of Percent cost
Transportation Percentage backhaul Total cost
backhauls savings of Computation
arc cost revenue routes matched savings with
without backhauls with time (sec)
parameter parameter to collaboration
collaboration collaboration
collaborators

MODEL 1
30% 10 $4,279.20 $563.52 13.17% 0.047
$1.20 /
40% 10 $4,279.20 $848.16 19.82% 0.032
mile
50% 13 $4,279.20 $1,156.80 27.03% 0.031
30% 10 $4,992.40 $657.44 13.17% 0.047
$1.40 /
40% 10 $4,992.40 $989.52 19.82% 0.047
mile
50% 13 $4,992.40 $1,349.60 27.03% 0.062
30% 10 $5,705.60 $751.36 13.17% 0.031
$1.60 /
40% 10 $5,705.60 $1,130.88 19.82% 0.047
mile
50% 13 $5,705.60 $1,542.40 27.03% 0.031
30% 10 $6,418.80 $845.28 13.17% 0.031
$1.80 /
40% 10 $6,418.80 $1,272.24 19.82% 0.047
mile
50% 13 $6,418.80 $1,735.20 27.03% 0.032
30% 10 $7,132.00 $939.20 13.17% 0.047
$2.00 /
40% 10 $7,132.00 $1,413.60 19.82% 0.047
mile
50% 13 $7,132.00 $1,928.00 27.03% 0.047
MODEL 2
30% 11 $4,279.20 $706.32 16.51% 0.203
$1.20 /
40% 11 $4,279.20 $941.76 22.01% 0.203
mile
50% 12 $4,279.20 $1,177.20 27.51% 0.218
30% 11 $4,992.40 $824.04 16.51% 0.218
$1.40 /
40% 11 $4,992.40 $1,098.72 22.01% 0.203
mile
50% 11 $4,992.40 $1,373.40 27.51% 0.203
30% 11 $5,705.60 $941.76 16.51% 0.218
$1.60 /
40% 11 $5,705.60 $1,255.68 22.01% 0.218
mile
50% 11 $5,705.60 $1,569.60 27.51% 0.218
30% 11 $6,418.80 $1,059.48 16.51% 0.234
$1.80 /
40% 11 $6,418.80 $1,412.64 22.01% 0.218
mile
50% 11 $6,418.80 $1,765.80 27.51% 0.188
30% 11 $7,132.00 $1,177.20 16.51% 0.218
$2.00 /
40% 11 $7,132.00 $1,569.60 22.01% 0.187
mile
50% 11 $7,132.00 $1,962.00 27.51% 0.218
6
7

TRB 2011 Annual Meeting Paper revised from original submittal.

 Bailey, Unnikrishnan, and Lin 14

1
2
3
4 TABLE 3 Model 2 Results for Number of Collaborators Serviced on Each Route
$1.20/mile $1.40/mile $1.60/mile $1.80/mile $2.0/mile
Route 30 40 50 30 40 50 30 40 50 30 40 50 30 40 50
% % % % % % % % % % % % % % %
1 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2
2 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
3 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
4 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
5 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
6 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
7 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
8 1 1 1 1 1 1 1 1 1 1 1 1 0 1 1
9 3 3 3 3 3 3 3 3 3 2 3 3 2 3 3
10 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
11 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
12 2 1 1 2 2 1 1 2 1 1 2 2 1 1 2
13 0 1 1 0 0 1 0 0 1 0 0 0 1 0 0
14 1 2 2 1 1 2 1 1 1 1 1 1 1 1 1
15 2 2 2 2 2 2 1 2 2 1 2 2 1 2 2
5
6
7
8
9 TABLE 4 Results with Variations in Unloading and Loading Times
Number of
Unloading and Total cost of Percent cost
backhaul Total cost
loading time backhauls savings of
routes matched savings with
parameters without backhauls with
to collaboration*
(hours) collaboration* collaboration
collaborators
MODEL 1
0.5 10 $5,705.60 $751.36 13.17%
0.75 10 $5,705.60 $751.36 13.17%
1 10 $5,705.60 $751.36 13.17%
1.25 7 $5,705.60 $660.58 11.58%
1.5 7 $5,705.60 $649.28 11.38%
MODEL 2
0.5 11 $5,705.60 $941.76 16.51%
0.75 11 $5,705.60 $941.76 16.51%
1 11 $5,705.60 $930.48 16.31%
1.25 8 $5,705.60 $726.48 12.73%
1.5 8 $5,705.60 $726.48 12.73%
10 *All trials use a transportation arc cost of $1.60/mile and 30% revenue
11

TRB 2011 Annual Meeting Paper revised from original submittal.

 Bailey, Unnikrishnan, and Lin 15

1
2
3 TABLE 5 Results with Reduction of the Maximum Backhaul Time Parameter
Number of
Total cost of Percent cost
Maximum Percentage backhaul Total cost
backhauls savings of
backhaul trip revenue routes matched savings with
without backhauls with
length parameter to collaboration
collaboration* collaboration
collaborators
MODEL 1
30% 10 $5705.60 $751.36 13.17%
Original 40% 10 $5705.60 $1,130.88 19.82%
50% 13 $5705.60 $1,542.40 27.03%
30% 7 $5705.60 $660.48 11.58%
10%
40% 8 $5705.60 $986.88 17.30%
Reduction
50% 9 $5705.60 $1,314.40 23.04%
MODEL 2
30% 11 $5705.60 $941.76 16.51%
Original 40% 11 $5705.60 $1,255.68 22.01%
50% 11 $5705.60 $1,569.60 27.51%
30% 8 $5705.60 $755.52 13.24%
10%
40% 9 $5705.60 $1,007.36 17.66%
Reduction
50% 9 $5705.60 $1,314.40 23.04%
4 *All trials use a transportation arc cost of $1.60/mile
5
6 CONCLUSION
7 In this paper, two mathematical models are presented to solve the problem of a carrier of interest
8 wanting to collaborate with other carriers and/or shippers to reduce operating costs along its
9 backhaul routes. Model 1 is formulated as an integer program and matches the carrier of
10 interest’s backhaul origins to either collaborator nodes or the depot node, while maximizing cost
11 savings for the carrier of interests. Model 2 is formulated as a mixed integer program, minimizes
12 the carrier of interest’s transportation costs and routes the backhaul trucks to collaborator pick-up
13 and delivery nodes where profitable, then routes back to the depot node. One distinguishable
14 difference between the models is that Model 1 allows each backhaul truck to fulfill at most one
15 collaborator’s shipping requests, while odel 2 allows more than one to be fulfilled.
16 The numerical analysis of the models using our example network reveal backhaul cost
17 savings between 13% to 28%. The original empty hauls to the depot are non-profitable trips and
18 any savings for the carrier of interest is significant. Through computational runs, it was found
19 that the arc transportation cost does not affect the percent cost savings of the solutions. The
20 revenue percentage level, the percent transportation cost that the carrier of interest charges for
21 delivering a collaborator’s shipment, affects the results of the models. The higher the percent
22 revenue, the more likely backhaul routes are matched with collaborators. Model 2 results show
23 that the percent cost savings is greater than the Model 1 results. This is because Model 2 allows
24 for each backhaul truck to fulfill more than one delivery task.
25 Areas of possible future research for this backhaul problem lie in developing a profit sharing
26 mechanism for the carrier of interest and the collaborative carriers/shippers. This could allow
27 more insight on the collaborator’s incentives to form partnerships with the carrier of interest.

TRB 2011 Annual Meeting Paper revised from original submittal.

 Bailey, Unnikrishnan, and Lin 16

1 ACKNOWLEDGEMENTS
2 The authors would like to acknowledge the support of Varunraj Valsaraj and other
3 members of the logistics team of Arrowstream Logistics for helping us in obtaining data,
4 problem formulation and providing us with guidance through discussions about the topic and the
5 implications of the various assumptions. The third author would like to acknowledge the
6 National Science Council, Taiwan, R.O.C. for providing partial funding support under project
7 no. NSC 99-2410-H-006-055
8
9
10 REFERENCES
11 [1] National Rail Freight Infrastructure Capacity and Investment Study, Cambridge
12 Systematics, Inc., Prepared for the Association of American Railroads, September 2007.
13
14 [2] Hernández, S. and P. Srinivas. Less-than-Truckload Static Single-Carrier Collaboration
15 Problem. Transportation Research Board Annual Meeting 2010 Paper #10-2574. CD-
16 ROM. Transportation Research Board of the National Academies, Washington, D.C.,
17 2010.
18
19 [3] Chu, C.W. A Heuristic Algorithm for the Truckload and Less-than-Truckload Problem.
20 European Journal of Operational Research, Vol. 165, 2005, pp. 657-667.
21
22 [4] Ball, M.O., B.L. Golden, A.A. Assad, and L.D. Bodin. Planning for Truck Fleet Size in
23 the Presence of a Common-Carrier Option. Decision Sciences, Vol. 14, No.1, 1983, pp.
24 103-120.
25
26 [5] Song, J. and A. Regan. An Auction Based Collaborative Carrier Network, CD-ROM,
27 Transportation Research Board of the National Academies, Washington, D.C., 2004.
28
29 [6] Figliozzi, M.A. Analysis and Evaluation of Incentive Compatible Dynamic Mechanism
30 for Carrier Collaboration. Transportation Research Record: Journal of the
31 Transportation Research Board, No. 1996, Transportation Research Board of the
32 National Academies, Washington, D.C., 2006, pp. 34-40.
33
34 [7] Dai, B. and C. Haoxun. Mathematical Model and Solution Approach for Collaborative
35 Logistics in Less Than Truckload (LTL) Transportation. Computers and Industrial
36 Engineering, 2009, pp. 767-772.
37
38 [8] Berger, S. and C. Bierwirth. Solutions to the Request Reassignment Problem in
39 Collaborative Carrier Networks. Transportation Research Part E, Vol. 46, No. 5, 2010,
40 pp. 627-638.
41
42 [9] Ergun, Ö., G. Kuyzu, and M. Savelsburgh. Shipper Collaboration. Computers and
43 Operation Research, Vol. 34, No. 6, 2007, pp. 1551-1560.
44
45 [10] Özener, O. Ö. and Ö. Ergun. Allocating Costs in a Collaborative Transportation
46 Procurement Network. Transportation Science, Vol. 42, No. 2, May 2008, pp. 146-165.

TRB 2011 Annual Meeting Paper revised from original submittal.

 Bailey, Unnikrishnan, and Lin 17

1
2 [11] Ergun, Ö., G. Kuyzu, and M. Savelsbergh. Reducing Truckload Transportation Costs
3 Through Collaboration. Transportation Science, Vol. 14, No. 2, May 2007, pp. 206-221.
4
5 [12] Agarwal, R., Ö. Ergun, and J. Sokol. Network Design and Allocation Mechanism for
6 Carrier Alliances in Liner Shipping. Under revision for Operations Research, Available
7 at: http://www2.isye.gatech.edu/~oergun/publications/ErgunPaper5Alliance_OR.pdf,
8 Last accessed in July2010.
9
10 [13] Agarwal, R. and Ö. Ergun. Mechanism Design for a Multicommodity Flow Game in
11 Service Network Alliances. Operational Research Letters, Vol. 36, No. 5, 2008, pp. 520-
12 524.
13
14 [14] Houghtalen, L., Ö. Ergun, and J. Sokol. Designing Mechanisms for the Management of
15 Carrier Alliances. Submitted to Management Science. Available at:
16 http://www2.isye.gatech.edu/~oergun/publications/AirAlliance_MS.pdf , last accessed in
17 July 2010.
18
19 [15] Arcelus, F.J., H.A. Eiselt and E.Y.H. Lin. Minimizing Empty Hauls in Multi-day, Multi-
20 depot Trucking. International Transactions Operational Research, Vol. 5, No. 3, 1998,
21 pp. 201-210.
22
23 [16] Jordan, W. Truck Backhauling on Networks with Many Terminals. Transportation
24 Research Part B, Vol. 21B, No. 3, 1987, pp. 183-193.
25
26 [17] Goetschalckx , M., C. Jacon-Blecha. The vehicle routing problem with backhauls.
27 European Journal of Operations Research, Vol. 42, 1989, pp. 39-51.
28
29 [18] Toth, P., D. Vigo. A heuristic algorithm for the symmetric and asymmetric vehicle
30 routing problems with backhauls. European Journal of Operational Research, Vol. 113,
31 1999, pp. 528-543.
32
33 [19] Dumas, Y., J. Desrosiers, F. Soumis. The pick-up and delivery problem with time
34 windows. European Journal of Operational Research, Vol. 54, 1991, pp. 7-22.
35
36 [20] Nagy, G. and S. Salhi. Heuristic algorithms for single and multiple depot vehicle routing
37 problems with pickups and deliveries. European Journal of Operational Research, Vol.
38 162, 2005, pp. 126-141.
39
40 [21] Toth, P. and D. Vigo, editors. The Vehicle Routing Problem. Vol. 9, SIAM Monographs
41 on Discrete Mathematics and Applications, SIAM, Philadelphia, PA, 2001.
42
43 [22] Gendreau, M., A. Hertz and G. Laporte. A Tabu Search Heuristic for the Vehicle Routing
44 Problem. Management Science, Vol 40(10), 1991, pp 1276-1290.
45

TRB 2011 Annual Meeting Paper revised from original submittal.

 Bailey, Unnikrishnan, and Lin 18

1 [23] Taillard, É., P. Badeau, M. Gendreau, F. Guertin, J.-Y. Potvin. A tabu search heuristic for
2 the vehicle routing problem with soft time windows. Transportation Science, Vol. 31,
3 1997 170–186.
4
5 [24] Cordeau, J.F.,M.Gendreau, and G. Laporte. A tabu search heuristic for periodic and
6 multi-depot vehicle routing problems. Networks, Vol 30(2), 1998, pp. 105-119
7
8 [25] Glover, F. Tabu Search: A Tutorial. Interfaces, Vol. 20, Issue 4, 1990, pp. 74-94.
9
10 [26] Hours-of-Service Regulations. Federal Motor Carrier Safety Administration, U.S
11 Department of Transportation. http://www.fmcsa.dot.gov/rules-
12 regulations/topics/hos/index.htm. Accessed May 2010.
13
14

TRB 2011 Annual Meeting Paper revised from original submittal.