Transportation Research Part E 42 (2006) 409430

www.elsevier.com/locate/tre

Air cargo eet routing and timetable setting with multiple on-time demands
Shangyao Yan *, Shin-Chin Chen, Chia-Hung Chen
Department of Civil Engineering, National Central University, Chungli 32054, Taiwan, ROC Received 7 July 2004; received in revised form 3 January 2005; accepted 5 February 2005

Abstract In this research we combine airport selection, eet routing and timetable setting to develop an integrated scheduling model. The objective is to maximize operating prot, given the related operating constraints. The model is formulated as a mixed integer program that is characterized as NP-hard. We develop several heuristics, and incorporate the use a mathematical programming solver, to solve the problem. To evaluate the model and the solution algorithms, we perform a case study using real operating data from a major Taiwan airline. The results show that the model and the solution algorithms could be useful for actual operations. 2005 Elsevier Ltd. All rights reserved.
Keywords: Cargo; Airport selection; Fleet routing; Timetables; Timespace network

1. Introduction According to predictions made by the International Civil Aviation Organization (ICAO), after AD 2000, airlines in the Asian-Pacic area will continue to dominate the international air freight (cargo) market. In recent years, Taiwan, in the center of this region, has striven to develop itself as an Asian and Pacic air freight transportation hub. Moreover, air freight markets in this area
*

Corresponding author. Tel.: +886 3 422 7151x34141; fax: +886 3 425 2960. E-mail address: t320002@cc.ncu.edu.tw (S. Yan).

1366-5545/$ - see front matter 2005 Elsevier Ltd. All rights reserved. doi:10.1016/j.tre.2005.02.002

410

S. Yan et al. / Transportation Research Part E 42 (2006) 409430

have become more competitive than before, so under these circumstances, Taiwans cargo carriers have had to provide better service of less cost, in order to maintain their market competitiveness. Fleet routing and ight scheduling are essential to a carriers operations. The setting of a good ight schedule cannot only enhance operating performance, but can also improve aircraft usage. Most cargo carrying airlines in Taiwan currently utilize a trial-and-error experience-based method for ight routing and scheduling, with the objective of maximizing the system prot or their market share. First a set of airports is rst selected with experience, then a timetable is drafted and a set of projected time periods are made that meet the cargo transport demands (i.e., the multiple on-time demands). The drafted timetable and the ight schedules are then adjusted according to eet availability, approved ight quotas of airports/airport pairs, airport turn-around times, the aircraft balance at each station, and other related operating costs/passenger revenues. This process is iterated manually, without optimization, from a systemic perspective. After adjustments, the schedule is then checked for aircraft maintenance and crew scheduling viability, and minor revisions may be made. Such a trial-and-error method is neither ecient nor eective, especially for large service networks. Much research by the air industry as well as academics has already been devoted to eet routing and ight scheduling problems. For example, Abara (1989) developed an integer linear programming model, formulated as a multi-commodity network ow problem, with xed ight departure times for eet assignment. Dobson and Lederer (1993) developed a three level hierarchical process to study the competitive choice of ight schedules and airfares by airlines in a pure hub-and-spoke (with single hub) system. Hane et al. (1995) modied Abaras model so that it could solve daily aircraft routing and scheduling problems (DARSP) without departure time windows. Clarke et al. (1996), based on Hane et al.s basic model, tried to develop a eet assignment model which would take aircraft maintenance and crew scheduling into consideration. Yan and Young (1996) developed a set of network models that could help a carrier eectively design short term ight schedules and eet routs given a drafted timetable and other operating constraints. Their models had the advantage of being more systematic and ecient than the traditional trial-and-error method. Desaulniers et al. (1997) proposed two integer programming models, a set partitioning type model and a time constrained multi-commodity network ow model, which could solve DARSPs utilizing a set of operational ight legs with known departure time windows. To improve Yan and Youngs model, Yan and Tseng (2002) developed an integrated scheduling model for multi-eet routing and ight scheduling. The objective was to maximize the system prot, given a xed projected passenger demand and other operating constraints. They also developed a Lagrangian relaxation-based algorithm to eciently solve the model. Barnhart et al. (2002) considered airline eet assignment problems involving prot maximization and the assignment of dierent aircraft types to dierent ight legs. They proposed a new formulation and solution approach that better captured network eects and generated superior solutions. Lohatepanont and Barnhart (2004) focused their attention on the steps of the airline schedule planning process that involved schedule design and eet assignment. They described integrated models and solution algorithms that could simultaneously optimize the selection of ight legs and the assignment of aircraft types to the selected ight legs. The above-mentioned research on ight scheduling has mainly focused on passenger transportation, which is fundamentally dierent from cargo transportation. In particular, the selection of airports in a passenger service network usually involves long-term planning, but in cargo trans-

S. Yan et al. / Transportation Research Part E 42 (2006) 409430

411

port, especially for Taiwans air carriers, this is not the case. To respond to signicant rapid uctuations in demand, carriers must perform their airport selection, eet routing and timetable setting to formulate short-term plans, while still considering demand and prot. Moreover, passengers are more time sensitive than cargos. Too many transfers in a passenger service may result in a signicant loss of customers, but cargos are not lost, provided they can be delivered on time. Research on freight transportation and eet routing has been performed by many researchers. For example, Chan and Ponder (1979) reviewed the air freight industry with special reference to the Federal Express Corporation. They outlined the characteristics of the industry and presented a survey of dierent managerial practices. Chestler (1985) described the basic structure of a pure hub-and-spoke network for air express carriers. Current et al. (1986) was the rst to introduce hierarchical network design problems. In subsequent research, Current et al. (1988) extended the design problems to include transshipment facilities with xed costs at the intersections of primary and secondary routes. Aykin (1995) addressed hub location and routing problems involving the joint determination of the location of interacting hubs and the types of service routes between these points. They presented a mathematical formulation of the problem and a solution algorithm. Jaillet et al. (1996) introduced three integer linear programming models for designing capacitated networks and routing policies, without the assumption of a hub-and-spoke structure. In the literature we found typically, that airport selection and frequency planning were considered in a service network design. Both are related to long-term planning. Our research, however, focuses in integrating airport selection, eet routing and timetable together in short-term operations. We note that, meta-heuristics have recently been employed to solve routing/scheduling problems, which are usually formulated as combinatorial optimization problems (e.g. see Gu and Huang, 1994 or Brandao and Mercer, 1997). Only a few meta-heuristics have been applied to solve pure (single commodity) network ow problems similar to the multi-commodity network ow problem discussed below. For example, Palmer and Kershenbaum (1995) developed a genetic algorithm (GA) that used Link and Node Biased (LNB) encoding method to solve an optimal communication spanning tree problem. Abuali et al. (1995) developed a GA that used a determinant encoding method to solve a probabilistic minimum spanning tree. Taguhi et al. (1998) presented a GA, with a non-uniform mutation and an arithmetic crossover, that could solve optimal ow assignment problems in computer networks. Yan and Luo (1998, 1999) employed the tabu search (TS), threshold accepting (TA), and simulated annealing (SA) methods in their development of several advanced local search algorithms that could be used to solve bipartite transportation network problems. Yan et al. (in press) utilized a GA to develop a global search algorithm for solving concave cost transshipment problems. However, the aforementioned meta-heuristics, which were all developed for solving pure network ow problems without side constraints, are difcult to apply to multi-commodity network ow problems. In this research, on the basis of the carriers perspective, we develop an integrated scheduling model that combines airport selection, eet routing and timetable setting, with the objective of maximizing the operating prot, given a set of projected cargo demands and the related operating constraints. We also develop several heuristic algorithms to nd solutions. Some airlines have already introduced combi ights into their operations, meaning passengers and cargo are transported at the same time. However, the airline used as an example in this research has only one

412

S. Yan et al. / Transportation Research Part E 42 (2006) 409430

type of cargo aircraft (B747-400F), so for simplicity, we have focused on pure cargo ight transportation to construct the model. It is expected that this model will be a useful planning tool for cargo carriers to determine suitable service airports, eet routes and timetables for short-term operations with on-time demands. The scope of this research is conned to cargo eet routing, ight scheduling and airport selection. Although, in practice, the related aircraft maintenance and crew scheduling processes must be considered during scheduling, these processes are usually modeled separately to facilitate problem solving (Teodorovic, 1986). For the studied Taiwan airline, the aircraft maintenance and crew constraints are actually rather exible, due to the use of stand-by crews and a progressive maintenance policy. These processes are performed after the eet routes and ights schedules have been solved. Thus, to reduce problem complexity, as done in conventional research, e.g. Yan and Young (1996) and Yan and Tseng (2002), we exclude these constraints in the modeling. The rest of this paper is organized as follows: rst, we introduce the model and develop several solution algorithms to solve the proposed model. Then, a case study is conducted to evaluate the performance of the model and the solution algorithms. Finally, some conclusions are oered.

2. Modeling approach A timespace network technique is applied to construct an integrated scheduling model that combines airport selection, eet routing and timetable setting, with the objective of maximizing the operating prot. The major elements in the modeling, including the eet-ow timespace network, the multi-cargo-ow timespace networks, and the mathematical formulation, are described below. 2.1. The eet-ow timespace network A timespace network, shown in Fig. 1, is established for single-eet routing within a specied time period (one week in this study) and specied locations. The horizontal axis represents the airport locations; the vertical axis stands for the time duration. All available airports are included. Nodes and arcs are the two major components in the network. Each node designates a specic airport and a specic time, while each arc represents an activity for an airplane, such as a ight leg, a ground holding period, or an overnight stay. The arc ows express the ow of airplanes in the network. Three types of arcs are dened below. 2.1.1. Flight leg arc A ight leg arc represents a ight connecting two airports. All possible ight legs between the two airports in the network, within a reasonable block of time, are considered, as long as time slots at the respective airports are available. Each ight leg arc contains information about the departure time, the departure airport, the arrival time, the arrival airport, and the operating cost. The time block for a ight leg is calculated as from the time when the airplane is prepared for this ight leg to the time when this ight leg is nished. Basically, this time includes the time for investigation prior to departure, fuelling, cargo loading and unloading, and ight time. The ight cost is the arc cost. The arc ows upper bound is one, meaning that the ight leg can be served at most

S. Yan et al. / Transportation Research Part E 42 (2006) 409430

Station - 1 Station - 2 Station - 3 Station - k
4:00 (day1) 8:00 (day1) 12:00 (day1) 16:00 (day1) 20:00 (day1)

413

(1)

(2)

24:00 (day1)

(3)

20:00 (day7) 24:00 (day7)

Fig. 1. Fleet-ow timespace network.

once. The arc ows lower bound is zero, implying that no airplane serves this ight leg. In addition, the departure interval at an airport is adjustable to meet the carriers operating requirements. Moreover, if a ight leg is served, then the two airports associated with that ight leg are used. A xed cost, as well as a variable cost for using the two airports, should be imposed. 2.1.2. Ground arc A ground arc represents the holding or the overnight stay of an aircraft at an airport in a time window. The arc cost, which includes the airport tax, the holding (or overnight stay) fee, the gate use charge and other related costs, denotes the expenses incurred for holding an aircraft at an airport in the corresponding time window. The arc ows upper bound is the apron capacity (or innity, if the capacity is large), indicating the maximum number of airplanes that can be held at this airport during a specic time window. The arc ows lower bound is zero, implying that no airplane is held at this airport in this time window. 2.1.3. Cycle arc A cycle arc represents the continuity between two consecutive planning periods. It connects the end of one period to the beginning of the next period for each airport. The arc cost is the cost of holding an aircraft overnight, and is similar to the ground arc cost, but with an additional overnight charge. The upper bound and lower bound of the arc ow are set as the same as those of the ground arcs.

414

S. Yan et al. / Transportation Research Part E 42 (2006) 409430

2.2. The cargo-ow timespace network The timespace network technique is also applied to formulate cargo movement corresponding to specied times (one week in this study) and locations, as shown in Fig. 2. Each cargo-ow time space network represents a specic OD pair from the origindestination table (known as the OD table). Such networks are designed to correspond to the eet-ow timespace network, so as to facilitate problem solving. In addition, it may be necessary for cargos with the same or dierent ODs to be delivered within dierent lengths of time, which means that the time window for delivery for an OD pair may be less than the length of the airline schedule (i.e., a week in this study). According to the time sensitivity of the cargos handled by the airline, we divide the time windows into three types (one day, four days, and one week). The horizontal and vertical axes are dened to be the same as those in the eet-ow timespace network. A node, here, also represents an airport at a specic time; however, an arc designates a cargo movement activity. Altogether, there are three types of arcs dened.

Station -1

Station -2

Station -3

Station -k 0 :00 (1 ) 4 :00 (1 ) 8 :00 (1 ) 12 :00 (1) 0 :00 (1) 4 :00 (1) 8 :00 (1) 12 :00 (1 ) 16 :00 (1 ) 20 :00 (1) 24 :00 (1) 4:00 (2 ) 8 :00 (2) 12 :00 (2) 16 :00 (2 ) 20 :00 (2) 24 :00 (2) 4 :00 (3) 8 :00 (3) : : : : : : 16 :00 (7) 20 :00 (7) 24 :00 (7 )

(3)

(2)

(1)

16 :00 (1) 20 :00 (1 ) 24 :00 (1 ) 4 :00 (2) 8 :00 (2) 12 :00 (2 ) 16 :00 (2) 20 :00 (2 )

(1) delivery arc (2) holding arc (3) demand arc 24 :00 (2 ) OD -time -pair : ( 1,2 ) ---- ( the 1 st layer )

OD -time -pair : ( 1,3 ) ---- ( the 2rd layer) OD -time -pair : ( m,m ) ---- (the nth layer)

Fig. 2. Cargo-ow timespace network.

S. Yan et al. / Transportation Research Part E 42 (2006) 409430

415

2.2.1. Delivery arc A delivery arc represents the transportation of the cargo from one airport to another on a specic ight. The transportation time is the same as the corresponding time block in the eet-ow timespace network for the associated ight. The arc cost is a variable cost for handling the cargo, per unit weight, on that ight, which is, in general, very small compared to the ight cost. The arc ows upper bound is the aircrafts capacity (denoted as the weight capacity for the studied airline). The arc ows lower bound is zero, indicating that no cargo from the corresponding OD pair is delivered on the associated ight. 2.2.2. Holding arc A holding arc indicates the holding of a cargo in a specic time window. The arc cost is the holding cost (or penalty) for this time window. However, if the arc just happens to connect either the departure or the arrival station of the corresponding OD pair, the arc cost is then zero, because whether a cargo is held before or after the delivery is usually not decided by the airline. Nevertheless, in practice the arc cost is adjustable. That is, if, in some special cases, a holding period at the O/D airport needs to be considered, then a suitable holding cost can be imposed. The arc ows upper bound is the stations cargo service capacity (or innity, if the capacity is relatively large), meaning the maximum amount of cargo (in weight units appropriate for the studied airline), that can be accommodated at this airport. The arc ows lower bound is zero, showing that no cargo from the corresponding OD pair is held at the airport during this time window. 2.2.3. Demand arc A demand arc, associated with an OD pair, connects the arrival station at the last time to the departure station at the rst time. It denotes the actual service demand for this OD pair. The arc cost is a negative value for cargo per unit weight delivered.1 The arc ows upper bound is the projected demand for this OD pair. The aim is to maximize prot, which implies that, for this OD pair, not all cargos will necessarily be served. The arc ows lower bound is zero, meaning that none are served. The trip demand for a specic OD pair can be exibly divided into several transportation time intervals, according to the actual delivery requirements. For example, the time length of the corresponding cargo-ow timespace network could be shorter for an OD pair including express deliveries. In contrast, the time length could be longer for a cargo that is less time sensitive. The time lengths are adjustable. 2.3. Notation and symbols used in the model Before introducing the model formulation, we list the notation and symbols used: the arc(i, j) ow in the eet network Xij the arc(i, j) ow in the nth cargo network Yn ij a decision variable, which equals 1 if station i is served, and 0 otherwise Wi
The cargo rate structure that an airline charges a forwarder or shipper is, in general, in decreasing ladder form (i.e., a concave function in terms of cargo accumulation). However, the amount of cargo transported on a ight is far larger than the charge to an individual customer. Consequently, in the planning scheduling stage, an average cargo fare per unit weight is usually used.
1

416

S. Yan et al. / Transportation Research Part E 42 (2006) 409430

the arc(i, j) cost in the eet network the arc(i, j) cost in the nth cargo network a xed cost for choosing station i, either as a departure airport or an arrival airport a variable cost at station i to handle cargos per unit weight, including loading and unloading n, N the nth OD pair and the set of all ODs A, NF the set of all arcs and nodes in the eet network Bn, NPn the set of all arcs and nodes in the nth cargo network AF the number of available airplanes in the eet network FF the set of all ight leg arcs in the eet network CF the set of all cycle arcs in the eet network BFn the set of all demand arcs in the nth cargo network the set of ight arcs associated with the ath station Sa the set of ight arcs that connect the ath station to the bth station Sab the approved ight quota at the ath station Qa ab the approved ight quota that connects the ath station to the bth station Q K the aircraft capacity (including a planning load factor) SA the set of all stations SP the set of airport pairs with an approved ight quota the arc(i, j) ows upper bound in the eet network Uij the arc(i, j) ows upper bound in the nth cargo network Un ij B a very large value

Cij Tn ij Fi Vi

2.4. The model formulation The model, given the eet-ow and the cargo-ow time space networks introduced above, is formulated as a mixed integer program. The objective of this model is to ow all aircraft and cargos simultaneously, in all networks at a minimum cost considering the cargo handling cost at the selected airports. Since the revenue from the cargo-ow networks is in the form of a negative cost, this objective is equivalent to the maximization of prot. There are several other issues that need to be carefully considered, such as: the number of aircraft required should not exceed the number of available airplanes; the accumulation of ights for a certain period at a specic airport/airport pair should not exceed the available quota; the amount of cargo carried on a ight should not exceed the capacity of that aircraft; and the airport selected should be able to handle cargo. Consequently, four types of side constraints are necessary for problem formulation: (1) the sum of the cycle arc ows in the eet-ow network should not be greater than the number of available airplanes; (2) the sum of all individual ights for each airport/airport pair should not exceed its approved ight quota; (3) the sum of all delivery arc ows corresponding to the same ight should not exceed the ight arc ow multiplied by the airplane capacity; and (4) the sum of all cargo network arc ows corresponding to an origin/destination airport should be less than or equal to a very large value times the binary variable value (either one or zero) associated with this airport. The model is formulated as follows:

S. Yan et al. / Transportation Research Part E 42 (2006) 409430

417

Min s:t::

Z X
j2NF

X
ij2A

C ij xij X
k 2NF

X X
n2N ij2Bn

n Tn ij y ij

X
i2SA

F i wi

X X
n2N ij2BFn

yn ij V i V j

1 2 3 4

xij yn ij

xki 0 yn ki 0

8i 2 NF 8i 2 NPn; 8n 2 N

j2NPn

k 2NPn

xij 6 AF xij 6 Qa xij 6 Qab yn ij 6 Kxij X X

n2N ij2BFn

ij2CF

8a 2 SA 8ab 2 SP 8ij 2 FF yn ij 6 Bwi 8i 2 SA

5 6 7 8 9 10 11 12 13

ij2S a

ij2S ab

X
n2N

06

8ij 2 FF xij 6 1 0 6 xij 6 U ij 8ij 2 A n n 0 6 y ij 6 U ij 8ij 2 Bn; 8n 2 N 8ij 2 A xij 2 I wi 2 f0; 1g 8i 2 SA

The model is formulated as a mixed integer multiple commodity network ow problem, in which the objective is to minimize the system cost. Constraints (2) and (3) ensure ow conservation at every node in each eet/cargo network. Eq. (4) denotes that the number of airplanes used in the eet network should not exceed the available number of airplanes. Eq. (5) ensures that the sum of all ights at each station does not exceed its approved quota. Eq. (6) ensures that the sum of all ight arcs connecting the ath station to the bth station does not exceed the approved ight quota. Eq. (7) keeps the cargo delivery volume within the aircrafts carrying capacity. Eq. (8) is used to determine whether a station is used for cargo serving or not. That is, if station i is used for serving cargos, then Wi = 1; otherwise Wi = 0. Eq. (9) indicates that each ight is served at most once. Eqs. (10) and (11) hold all the arc ows within their bounds. Eq. (12) ensures the integrality of the airplane ows. Eq. (13) indicates that each airport selection decision is binary.2

3. Solution method The model is formulated as a mixed integer program that is characterized as NP-hard (Garey and Johnson, 1979). It is almost impossible to optimally solve a realistically large problem within
If a carrier has to provide a minimum number of ights for certain station-pairs, due to a specic marketing strategy, then a constraint can be introduced for each of the station-pairs, to ensure that the sum of the associated ight leg arc ows is greater than or equal to the minimum number of ights.
2

418
origin

S. Yan et al. / Transportation Research Part E 42 (2006) 409430

destination origin destination origin destination

(a) Non-stop network

(b) One-stop network

(c) All-stop network

Fig. 3. Heuristic Networks.

a limited time. For example, in numerical tests the all-stop heuristic (Section 4.1) took 110668.77 s to solve a small-scale problem with only eight stations and four airplanes. Therefore, we must develop a family of heuristics to solve large-scale problems. The heuristics take into account the stops required by the cargo to be transported, as described below: (1) Non-stop heuristic: as shown in Fig. 3a, all cargos are delivered using non-stop ights that go directly from their point of origin to their destination. In this heuristic, the delivery arcs in each cargo network need only be set from their associated origin to their associated destination. Other delivery arcs are removed from the cargo network. The modied model, the nonstop network, can then be optimally solved. (2) One-stop heuristic: as shown in Fig. 3b, all cargos are transported using non-stop and/or one-stop ights. In this heuristic, the delivery arcs in each cargo network need only be set from the associated origin to all other stations and from all other stations to the associated destination. Other delivery arcs are removed from the cargo network.3 The modied model, the one-stop network, can then be optimally solved. (3) All-stop heuristic: as shown in Fig. 3c, all cargos are transported using non-stop, one-stop, and/or multiple stop ights. In this heuristic, all possible delivery arcs between two airports in each cargo network are set.4 The modied model, the all-stop network, can then be optimally solved. Obviously, the problem size of the all-stop network is larger than that of the one-stop network, which is larger than that of the non-stop network. (4) Mixed-stop heuristic: Cargos are transported using non-stop, one-stop, and/or multiple stop ights, according to the OD distance. Although the all-stop network is useful for nding the most eective transport plan, its problem scale may be too large to solve. To suitably reduce the problem size of the all-stop network, the delivery arcs in this heuristic, in each cargo network, are set according to their OD distance. In particular, in accordance with real practice,
There is no delivery arc pointing to the associated origin or emanating from the associated destination. Similar in the one-stop network, there is no delivery arc pointing to the associated origin or emanating from the associated destination.
4 3

S. Yan et al. / Transportation Research Part E 42 (2006) 409430

Start

419

Solve the modified problem, and compare the objective value with the lower bound

Solve all-stop networks using CPLEX Converges?

Yes

Yes

Converges?
No

No

Check all cargo networks with an OD distance greater than L1 have been modified into onestop networks

Set the lower bound solution obtained from CPLEX Satisfied? Set L1 and L2 to represent boundary distances
Yes

Modify the selected networks from one-stop networks to allstop networks.

Solve the non-stop network using CPLEX, and compare the objective value with the lower bound

No

Solve the modified problem, and compare the objective value with the lower bound

Yes

Converges?

No

Converges?
No

Yes

No

Modify the selected networks from non-stop networks to one-stop networks

Check that all cargo networks with an OD distance greater than L2 have been modified into all-stop networks

Satisfied?
Yes

Use the best solution obtained above as the final solution End

Fig. 4. Flowchart of the solution method.

short-haul OD cargo networks are designated as non-stop networks. Similarly middle-haul/ long-haul OD cargo networks are designated as one-stop/all-stop networks. The modied model, the mixed-stop network, can then be optimally solved. This model can eectively provide multi-stop/non-stop ights for long-haul OD cargos, one-stop/non-stop ights for middle-haul OD cargos, and non-stop ights for short-haul OD cargos.

420

S. Yan et al. / Transportation Research Part E 42 (2006) 409430

(5) Improved mixed-stop heuristic: Unlike the four heuristics that use the mathematical programming solver, CPLEX, to directly solve the associated networks, in this heuristic, we repeatedly modify and solve a series of networks using CPLEX, to approximate a nearoptimal solution, an improvement on the mixed-stop heuristic. The owchart of the solution method is shown in Fig. 4, and the steps are listed as follows: Step 1: Solve the all-stop network in a pre-set time using CPLEX. If the solution is occurring within the specied error tolerance, then stop. Otherwise, set the lower bound solution obtained from CPLEX as the lower bound and go to step 2. Step 2: Solve the non-stop network using CPLEX, then compare the objective value with the lower bound. If it falls within the specied error tolerance, then stop. Otherwise, go to step 4. Step 3: Set L1 and L2 to represent the respective boundary distances dening short-haul and middle-haul ights, as well as middle-haul and long-haul ights (L1 < L2). Step 4: Sort the cargo networks according to increasing OD distances. Choose a number p1 of non-stop cargo networks, sequentially from the top, with OD distances greater than L1. The p1 value which results in good solution quality can be tested. In practice, a carrier may nd a number of p1 values, as in this research, to be tested. Modify the networks selected from the non-stop networks to be one-stop networks. Solve the modied problem using CPLEX and compare the objective value with the lower bound. If it falls within the specied error tolerance, then stop. If all cargo networks with an OD distance greater than L1 have been modied into one-stop networks, then go to step 5. Otherwise, return to step 4. Step 5: Sort the cargo networks according to increasing OD distance. Choose a number p2 of one-stop cargo networks, sequentially from the top. Their OD distance should be greater than L2.5 Modify the networks selected from the one-stop networks to be all-stop networks. Solve the modied problem using CPLEX and compare the objective value with the lower bound. If it falls within the specied error tolerance, then stop. If all cargo networks with an OD distance greater than L2 have been modied into all-stop networks, then go to step 6. Otherwise, return to step 5. Step 6: Use the best solution obtained above as the nal solution. 4. Numerical tests To test how well the model and the solution algorithms may be applied in the real world, we performed a case study using operational data from a major Taiwan airline. To build and to solve the models, we used the C computer language, coupled with the CPLEX 7.1 mathematical programming solver, to develop the solution algorithms. The tests were performed on a Pentium 4-1.8G with 1 Gb of RAM in the environment of Microsoft Windows 2000. We rst used the operating data to build a mathematical model, and then applied the solution algorithms to solve the problems. Finally, we performed several sensitivity/scenario analyses.
5

Similar to the setting of p1, the suitable p2 value should be tested.

S. Yan et al. / Transportation Research Part E 42 (2006) 409430

421

4.1. Model tests and result analyses The numerical tests were mainly based on data obtained from a major Taiwan airlines Southeast Asian operations during 2002. Eight cities were served by four B747-400F airplanes, each with a carrying capacity of 100 metric tons. All the cost parameters, cargo fare rates and other input, such as the ight times, the distance between two stations, the approved ight quota for each airport/airport pair, the available time slots at each airport, and the ground handling time, were set based on the actual operating data, as well as Taiwan government regulations, with reasonable simplications. The heuristics developed included non-stop, one-stop, all-stop, mixed-stop and improved mixed-stop heuristics. After testing to nd suitable L1 and L2 values that would result in good solution quality, L1 and L2 were set to be 1500 km and 3000 km.6 In other words, ights with OD distances of less than 1500 km, between 1500 km and 3000 km, or greater than 3000 km would be characterized as short-haul, middle-haul or long-haul ights, respectively. In addition, in step 4 of the improved mixed-stop heuristic, we chose three non-stop cargo networks (with OD distances greater than 1500 km) in each round, to modify into one-stop networks. In step 5 of the improved mixed-stop heuristic, we chose one one-stop cargo network (with an OD distance greater than 3000 km) in each round, to modify into an all-stop network. The problem sizes handled by each heuristic, with the exception of the improved mixed-stop heuristic, are described in Table 1.7 In the tests, we set the convergence gap (error tolerance) to be 0.05 and used CPLEX to solve the problems. All the problems were solved to within an error gap of 0.05. To evaluate the performance of the heuristics, we utilized the current timetable (24 non-stop ights and 4 one-stop ights) and eet routing (4 airplanes) of the studied airline with the projected cargo transportation demand. In particular, we rst xed the values of the eet-ow timespace network variables, given the timetable and eet routing. Then we used CPLEX to solve a simplied cargo transportation model. For simplicity, the results obtained are referred to as actual operations. Table 2 shows the test results. As shown in Table 2, OBJ represents the system cost for the best feasible solution obtained, which is equal to a negative prot, because the revenue from the cargoow networks is calculated in the form of a negative cost. For example, the all-stop heuristic, as shown in Table 2, has a system cost of 26,413,886 which is equal to a prot of 26,413,886. For simplicity, we use the system cost to later interpret the test results. Best Node represents the best objective function value of all the unexplored nodes in the branch-and-bound tree, and serves as the lower bound of the problem. Gap represents the gap between OBJ and Best Node. The improved mixed-stop heuristic yielded the best solution, with an objective value of 26,712,087. The mixed-stop heuristic was next, with an objective value of 26,510,486, followed by the one-stop heuristic, with an objective value of 26,472,688. The all-stop heuristic performed slightly worse than the one-stop heuristic, with an objective value of 26,413,886, followed by the actual operation, with an objective value of 24,433,012. The non-stop heuristic performed most
The search for the most suitable L1 and L2 is performed similar to that for nding the most suitable p value; See step 4 in the improved mixed-stop heuristic. L1 and L2 are adjustable in other applications. 7 The problem size handled by the improved mixed-stop heuristic changes in each round.
6

422 Table 1 Problem size Heuristic # Variables # Constraints

S. Yan et al. / Transportation Research Part E 42 (2006) 409430

Non-stop 5142 4757

One-stop 19,614 9737

All-stop 43,734 9737

Mixed-stop 25,104 7841

Table 2 Test results Heuristic OBJ (NT$) Best Node (NT$) Gap (%) Number of iterations for running CPLEX Computation time (s) Number of stations Fleet size Frequency (ights/week) Average load factor (%) Transfer rate (%) Service rate (%) One-stop ight Two-stop ight Actual operations Non-stop 24,433,012 25,031,259 2.39% 859 91.43 8 4 28 67.18% 3.32% 99.15% 4 0 2,402,5318 24,963,600 3.91% 30 4.92 8 3 29 65.52% N/A 99.52% N/A N/A One-stop 26,472,688 27,675,212 4.54% 14,578 3072.16 8 3 28 68.18% 9.52% 100.00% 6 0 All-stop 26,413,886 27,734,232 5.00% 15,200 110668.77 8 3 28 67.86% 5.74% 99.52% 4 1 Mixed-stop 26,510,486 27,680,854 4.41% 26,788 39748.61 8 3 28 68.18% 9.52% 100.00% 6 0 Improved mixed-stop 26,712,087 27,869,619 4.33% 13,365 16843.33 8 3 28 68.18% 9.52% 100.00% 6 0

poorly, with an objective value of 2,402,5318. The results show that a mixture of non-stop ights, one-stop ights and multiple-stop ights, given by the improved mixed-stop heuristic, provided the most protable schedule, while the purely non-stop ights given by the non-stop heuristic will yielded the least protable schedule. The results also show that, except for the non-stop heuristic, the objective value of the other four heuristics were all better than actual operating value by at least 8.03%, which shows that, from a systematic optimization perspective the proposed heuristics are superior to the current trial-and-error method. Although the non-stop heuristic is designed for optimization analysis, it yielded an objective value that was worse than the actual operating value. This may be a result of the fact that the non-stop heuristic does not take into account one-stop ight operations, which are relatively non-systemic compared to the trial-and-error method currently used, in the Southeast Asia network.8 It is also found that except for the non-stop heuristic, the other four heuristics all yielded 46 one-stop ights. The all-stop heuristic also yielded a two-stop ight. The objective values of these four heuristics are close to each other, the greatest dierence being 1.1%. The reason is that the
According to the optimality theory, the more exible the model, the better the solution obtained. Consequently, the all-stop network should provide the best solution, if all networks are optimally solved. However, due to the 5% error gap set for convergence in CPLEX, in the tests the all-stop heuristic did not yield better solutions than the improved mixed-stop heuristic, the mixed-stop heuristic or the one-stop heuristic.
8

S. Yan et al. / Transportation Research Part E 42 (2006) 409430

423

test was based on the carriers Southeast Asia service network, where most distances between airports are short. Therefore, a combination of non-stop and one-stop ights would be most eective for most cargos. More than one-stop ights are not signicantly eective in this area, which is why only one two-stop ight is employed in the all-stop heuristic solution. If a larger ight network needs to be tested, then more multiple-stop ights should be used. The all-stop heuristic used more computation time than did the others, since it dealt with the most complicated network. For example, as shown in Table 1, the number of variables in the all-stop network was 2.22 times that in the one-stop network. The numbers of constraints in both these networks were the same. However, the computation time required by the all-stop network was 36.02 times of that required by the one-stop network, showing that for NP-hard problems, the computation time will increase substantially when the number of variables increases. Although the all-stop network incorporated the most information, its objective was not the best, due to the convergence gap set in CPLEX. A similar result was also found for the mixed-stop heuristic. Of the ve heuristics, the improved mixed-stop heuristic not only had a better objective value than the other heuristics, but was also less time-consuming than the all-stop heuristic and the mixed-stop heuristic (requiring only 15% of the all-stop heuristics and only 42% of the mixed-stop heuristics computation time). As a result, the improved mixed-stop heuristic was superior to the others in terms of solution quality and computational eciency. In the results, 8 stations and 3 airplanes were used for all ve heuristics. The non-stop heuristic provided 29 ights, while the other four heuristics provided 28 ights. In addition, the average load factor for the non-stop heuristic was 65.52%, which was lower than for the other four heuristics and for actual operations. We also found that eective cargo transfers not only led to an improvement in the average load factor but also the cargo service rate. For example, all the one-stop, mixed-stop and improved mixed-stop heuristics yielded the highest transfer rate, 9.52%, the highest average load factor, 68.18% and the highest cargo service rate, 100%, meaning that the resources had been most eectively used.9 The other two heuristics (the non-stop and the all-stop heuristics) and actual operations yielded worse results, in particular lower transfer rates (none, 5.74% and 3.32%, respectively), lower load factors (65.52%, 67.86% and 67.18%, respectively) and lower cargo service rates (99.52%, 99.52% and 99.15%, respectively), meaning that resources were relatively less eectively used. Finally, the eet ows obtained above could not yet be directly put into practice without identifying each airplane path in the eet networks. The ow decomposition method (Yan and Young, 1996) was applied to trace the path of each airplane. An example of the three aircraft routes is shown in Fig. 5. 4.2. Sensitivity/scenario analyses To understand the inuence of the model parameters on the solution, we performed a sensitivity analysis of the available eet size, the cargo demand (OD), and the xed station cost all of
The results, except for the objective value, from the mixed-stop and the improved mixed-stop heuristics were all the same. We traced the routes and found that these two routes, except for the transfer times of some cargos, were the same. In particular, the cargo holding costs was what made the objective values be dierent.
9

424

S. Yan et al. / Transportation Research Part E 42 (2006) 409430

Fig. 5. Aircraft route example.

which are essential inputs to the model.10 For simplicity, we used the two better heuristics, that is, the mixed-stop and the improved mixed-stop heuristics, for the analyses. To evaluate the performance of the proposed heuristics for solving middle/large scale problems, we also performed a scenario analysis.

10

Sensitivity analyses of other factors can be similarly performed, but are left for future research.

S. Yan et al. / Transportation Research Part E 42 (2006) 409430

Mixed-stop heuristic Improved mixed-stop heuristic

425

-20000000

Objective value (NT$)

-22000000 -24000000 -26000000 -28000000

Available fleet size

Fig. 6. Sensitivity analysis of the available eet size.

4.2.1. Available eet size As shown in Fig. 6, the objective value of the mixed-stop heuristic decreased from 20,427,560 to 26,497,732 (i.e., a prot improvement of 29.72%), when the available eet size increased from 1 to 3 airplanes.11 However, when the available eet size increased from 3 to 4 airplanes, the optimal eet size remained at 3 airplanes. Although the objective value changed slightly, from 26,497,732 to 26,510,486 (a dierence of 0.05%), this was due to the convergence gap set in CPLEX. The situation was found to be similar for the improved mixed-stop heuristic. 4.2.2. Cargo demand To evaluate the potential for a future increase in cargo demand, we performed a cargo demand sensitivity analysis. As shown in Fig. 7, the objective value of the mixed-stop heuristic decreased when the cargo demand increased, however, the decrement tended to be smaller. For example, the objective value decreased from 26,510,486 to 30,813,412 (i.e., a prot improvement of 16.23%), when the cargo demand increased from 100% to 120%. However, when the cargo demand increased from 180% to 200%, the objective value decreased from 40,238,048 to 41,514,180 (i.e., a prot improvement of 3.17%). The latter is signicantly smaller than the former.12 Similar situations were also found for the improved mixed-stop heuristic. 4.2.3. Stations xed cost The xed cost for operating each station was estimated based on the airlines annual report. As shown in Fig. 8, the objective value of the mixed-stop heuristic increased when the xed cost increased, meaning that the carriers prot decreased. When every stations xed cost increased from 100% to 140%, the objective value increased from 26,510,486 to 26,171,030 (i.e., the prot decreased by 1.3%), meaning that the prot was relatively uninuenced by the stations xed cost. A similar situation was found for the improved mixed-stop heuristic, except its objective values were
All available airplanes were used up. The service rate decreased as the demand increased. This decrease was a result of the fact that the sum of all ights at each station exceeded the approved quota, so that additional cargos could not be serviced.
12 11

426

S. Yan et al. / Transportation Research Part E 42 (2006) 409430

Mixed-stop heuristic Improved mixed-stop heuristic

80% -25000000

100% 120% 140% 160% 180% 200% 220%

Objective value (NT$)

-30000000 -35000000 -40000000 -45000000

Cargo demand (%)

Fig. 7. Sensitivity analysis of the cargo demand.

Mixed-stop heuristic Improved mixed-stop heuristic

80% -24000000

100%

120%

140%

160%

Objective value (NT$)

-25000000 -26000000 -27000000 -28000000

Station fixed cost (%)

Fig. 8. Sensitivity analysis of the stations xed cost.

slightly better. Both analyses showed that the stations xed cost was not sensitive to the modeled objective value. 4.2.4. Problem scales To evaluate the performance of the proposed heuristics for middle/large scale problems, we further tested seven problem instances with dierent scales ranging from 10 cities, 10 airplanes and 30 OD pairs, to 24 cities, 30 airplanes and 60 OD pairs. For each problem instance, we added a number of cities, airplanes and OD pairs to our original problem. The distance and the ight times for each new city-pair were randomly set, based on the original ight network. The available time slots and the ground handling time at each new airport were randomly set, in reference to the original problem. Based on the original OD demands, as well as the original/new eet size, we randomly set the cargo demand for each OD pair. In particular, the cargo demand for each OD pair was suitably increased with the eet size. Similarly, the ight quota for every airport/airport pair was increased when the associated OD demand increased. The other cost parameters and cargo fare rates were randomly selected in reference to the original problem. The parameters used in

S. Yan et al. / Transportation Research Part E 42 (2006) 409430 Table 3 Test results for middle/large-scale problems Non-stop 10 Cities, 30 OD pairs and 10 airplanes # Variables 6508 # Constraints 5843 OBJ (NT$) 101,243,344 Best Node (NT$) 106,153,768 Gap (%) 4.85 Computation time (s) 4.53 10 Cities, 30 OD pairs and 15 airplanes # Variables 6508 # Constraints 5843 OBJ (NT$) 103,373,504 Best Node (NT$) 107,843,288 Gap (%) 4.32 Computation time (s) 5.28 10 Cities, 30 OD pairs and 20 airplanes # Variables 6508 # Constraints 5843 OBJ (NT$) 103,296,144 Best Node (NT$) 107,543,088 Gap (%) 4.11 Computation time (s) 7.15 16 Cities, 42 OD pairs and 10 airplanes # Variables 13,902 # Constraints 12,957 OBJ (NT$) 206,525,344 Best Node (NT$) 214,112,736 Gap (%) 3.67 Computation time (s) 5.85 16 Cities, 42 OD pairs and 15 airplanes # Variables 13,902 # Constraints 12,957 OBJ (NT$) 234,796,080 Best Node (NT$) 242,086,528 Gap (%) 3.11 Computation time (s) 7.12 16 Cities, 42 OD pairs and 20 airplanes # Variables 13,902 # Constraints 12,957 OBJ (NT$) 244,625,040 Best Node (NT$) 252,345,952 Gap (%) 3.16 Computation time (s) 8.36 One-stop All-stop Mixed-stop Improved mixed-stop 106,732,979 110,599,832 3.49 100.09

427

36,987 12,209 105,914,328 110,535,352 4.36 15.34

60,257 12,209 10,571,3709 110,613,301 4.63 2348.56

24,782 9219 106,533,160 110,380,315 3.61 354.16

36,987 12,209 108,920,480 112,651,880 3.43 21.73

60,257 12,209 107,632,826 112,973,599 4.96 1092.50

24,782 9219 107,982,959 112,287,897 3.99 607.98

109,419,370 112,675,698 2.89 207.31

36,987 12,209 107,785,200 112,701,832 4.56 25.37

60,257 12,209 109,031,509 112,838,800 3.49 1917.98

24,782 9219 107,982,959 112,287,897 3.99 645.97

109,109,504 112,658,239 3.15 201.76

85,480 28,119 211,680,131 222,049,859 4.67 153.48

222,075 28,119 211,849,793 222,064,772 4.60 55574.03

65,017 20,708 211,732,551 222,055,376 4.63 1066.63

211,890,851 222,061,256 4.58 324.47

85,480 28,119 244,945,456 253,776,892 3.48 128.55

222,075 28,119 245,245,480 253,797,154 3.49 22806.84

65,017 20,708 245,441,219 253,738,467 3.27 644.17

245,630,038 253,776,256 3.21 211.61

85,480 28,119 257,951,790 265,984,523 3.02 138.23

222,075 28,119 257,775,096 266,216,436 3.04 30560.70

65,017 20,708 257,645,875 265,553,795 3.07 2007.11

261,146,491 266,095,874 1.86 602.41

(continued on next page)

428 Table 3 (continued)

S. Yan et al. / Transportation Research Part E 42 (2006) 409430

Non-stop 24 Cities, 60 OD pairs and # Variables # Constraints OBJ (NT$) Best Node (NT$) Gap (%) Computation time (s) 30 airplanes 30,902 27,296 336,685,344 350,772,576 4.18 9.89

One-stop

All-stop

Mixed-stop

Improved mixed-stop 381,972,634 39,368,5142 2.97 1985.24

235,684 60,620 380,294,470 394,005,874 3.48 1358.72

584,260 60,620 ** 394,330,000 ** 259,807

151,326 42,241 381,628,506 393,701,862 3.16 6392.56

: The problem size handled by the improved mixed-stop heuristic changes in each round. **: The all-stop heuristic could not nd a feasible solution in three days (259,200 s).

the heuristics and CPLEX were set the same as in the previous test. Furthermore, the computational capability was upgraded. The tests were performed on a Pentium 4-3.2G with 1.5 GB of DDR RAM in the environment of Microsoft Windows XP. The CPLEX version was upgraded from 7.1 to 8.1. Table 3 shows the test results for dierent problem instances. The results similar to that of the original smaller problem, show that the proposed heuristics, except for the all-stop heuristic, could eciently solve middle/large-scale problems. For example, as shown in the 16 city, 42 OD pair and 20 airplane scenario, the improved mixed-stop heuristic yielded the best solution, with an objective value of 261,146,491. The non-stop heuristic performed the worst, with an objective value of 244,625,040. The other three heuristics, the mixed-stop heuristic, the all-stop heuristic and the one-stop heuristic, yielded similar objectives within 0.12% (= j(257,951,790 257,645,875)/ 257,951,790j of each other). However, the all-stop heuristic was signicantly more time-consuming. For example, the all-stop network required a computation time 15.22 times that of the mixedstop network and 50.73 times that of the improved mixed-stop network. The other problem results are similar and not discussed here.13

5. Conclusions In the past, most research on airline scheduling has been focused on passenger transport, which is fundamentally dierent from air cargo transport. A short-term scheduling model has not yet been developed for air cargo transportation. In this research, based on Taiwan air carrier data, we developed a novel integrated scheduling model combining airport selection, eet routing and timetable setting, with the objective of maximizing the operating prot, subject to the related operating constraints. Such a model is expected to be a useful planning tool, with which cargo carriers can determine suitable servicing airports, eet routes and timetables for short-term operations. Network ow techniques are employed to construct the model, which includes multiple cargo- and eet-ow networks. In the cargo-ow networks, unlike in past research, we construct
13

To save space, other results are not discussed here. The reader may contact the authors for the detail.

S. Yan et al. / Transportation Research Part E 42 (2006) 409430

429

multiple OD-time-pair timespace networks on the basis of the desired delivery times. The model is formulated as a mixed integer program that is characterized as NP-hard. We develop several heuristics with the use of a mathematical programming solver to solve the problem. To evaluate the model and the solution algorithms, a case study based on real operating data from a major Taiwan cargo carrier is performed. The results show that in terms of solution quality and computational eciency the improved mixed-stop heuristic out-performs the others. The results also show that the use of suitable cargo transfers not only reduces the operating cost but also increases protability. To understand how the essential parameters inuenced the solution, several sensitivity analyses are also performed. From the results, we see that the improved mixed-stop heuristic performs better than the mixed-stop heuristic, in each test. Additional testing of middle/large-scale problems is performed. The results, similar to that for the small problem, show that the proposed heuristics, except the all-stop heuristic, can all eciently solve middle/large-scale problems. Since those test problems are articial, practical problems should next be tested for the carrier to evaluate the proposed heuristics in actual operations. The model, the solution algorithms, the case study, and the sensitivity/scenario analyses should all be useful as reference material, to help an airline to determine optimal short-term cargo eet routing and ight scheduling. Although the preliminary test results show that the proposed heuristics, except the all-stop heuristic, have potential to be used for solving middle/large-scale problems, the heuristics can be suitably modied to solve larger-scale problems, for larger airlines. For example, if the heuristics as proposed cannot eciently solve large-scale problems, then modern meta-heuristic techniques (e.g. the tabu search method, threshold accepting method, or genetic algorithm), lagrangian relaxation or column generation, may be employed to develop a more ecient algorithm. This could be a direction of future research. The extension of single-eet routing to multi-eet routing, and the incorporation of other routing constraints (for example, maintenance and crew scheduling) or other objectives involved in actual operations, could also be directions for future research. Here, we used cargo ights for the input data. Some airlines have introduced combi ights. How to combine passenger ights and cargo ights into an integrated ight scheduling model could also be a topic of future research. Finally, the proposed model and solution techniques could be applied to other transport industries, for example, container ship scheduling or bus scheduling. Acknowledgments This research was supported by a grant (NSC-92-2211-E-008-046) from the National Science Council of Taiwan. We thank the airline for providing the test data and their valuable opinions. We also thank the two anonymous referees and the editor for their helpful comments and suggestions on the presentation of the paper.

References
Abara, J., 1989. Applying integer linear programming to the eet assignment problem. Interfaces 19, 2028. Abuali, F.N., Wainwright, R.L., Schoenefeld D.A., 1995. Determinant factorization: a new encoding scheme for spanning trees applied to the probabilistic minimum spanning tree problem. In: Proceedings of the Sixth International Conference on Genetic Algorithms, pp. 470477.

430

S. Yan et al. / Transportation Research Part E 42 (2006) 409430

Aykin, T., 1995. The Hub location and routing problem. European Journal of Operational Research 83, 200219. Barnhart, C., Kniker, T.S., Lohatepanont, M., 2002. Itinerary-based airline eet assignment. Transportation Science 36 (2), 199217. Brandao, J., Mercer, A., 1997. A tabu search algorithm for the multi-trip vehicle routing and scheduling problems. European Journal of Operational Research 100, 180191. Chan, Y., Ponder, R., 1979. The small package air freight industry in the united states: a review of the federal express experience. Transportation Research-A 13, 221229. Chestler, L., 1985. Overnight air express: spatial pattern, competition and the future of small package delivery services. Transportation Quarterly 39, 5971. Clarke, L.W., Hane, C.A., Johnson, E.L., Nemhauser, G.L., 1996. Maintenance and crew considerations in eet assignment. Transportation Science 30, 249260. Current, J.R., Revelle, C.S., Cohon, J.L., 1986. The hierarchical network design problem. European Journal of Operations Research 2, 5766. Current, J.R., Revelle, C.S., Cohon, J.L., 1988. The design of a hierarchical transportation network with transshipment facilities. Transportation Science 22 (4), 270277. Desaulniers, G., Desrosiers, J., Dumas, Y., Solomon, M.M., Soumis, F., 1997. Daily aircraft routing and scheduling. Management Science 43, 841855. Dobson, G., Lederer, P.J., 1993. Airline scheduling and routing in a hub-and-spoke system. Transportation Science 27 (3), 281297. Garey, M.R., Johnson, D.S., 1979. Computers and Intractability: a Guide to the Theory of NP-completeness. W.H. Freemean & Company, San Francisco. Gu, J., Huang, X., 1994. Ecient local search with search space smoothing: a case study of the traveling salesman problem (TSP). IEEE Transaction on Systems, Man and Cybernetics 24, 728739. Hane, C.A., Barnhart, C., Johnson, E.L., Marsten, R., Nemhauser, G.L., Sigismondi, G., 1995. The eet assignment problem: solving a large-scale integer program. Mathematical Programming Study 70, 211232. Jaillet, P., Gao, S., Yu, G.., 1996. Airline network design and hub location problem. Location Science 14 (3), 195212. Lohatepanont, M., Barnhart, C., 2004. Airline schedule planning: integrated models and algorithms for schedule design and eet assignment. Transportation Science 38 (1), 1932. Palmer, C.C., Kershenbaum, A., 1995. An approach to a problem in network design using genetic algorithms. Networks 26, 151163. Taguhi, T., Ida, K., Gen, M., 1998. A genetic algorithm for optimal ow assignment in computer network. Computers and Industrial Engineering 35 (34), 535538. Teodorovic, D., 1986. Multi-attribute aircraft choice for airline network. Journal of Transportation Engineering 112, 634646. Yan, S., Luo, S.C., 1998. A tabu search-based algorithm for concave cost transportation network problems. Journal of the Chinese Institute of Engineers 21, 327335. Yan, S., Luo, S.C., 1999. Probabilistic local search algorithms for concave cost transportation network problems. European Journal of Operational Research 117, 511521. Yan, S., Tseng, C.H., 2002. A passenger demand based model for airline ight scheduling and eet routing. Computers and Operations Research 29, 15591581. Yan, S., Young, H.F., 1996. A decision support framework for multi-eet routing and multi-stop ight scheduling. Transportation Research 30A, 379398. Yan, S., Juang, D.S., Chen, C.R., Lai, W.S., in press. Global and local search algorithms for concave cost transshipment problems. Journal of Global Optimization.