Vous êtes sur la page 1sur 15

Transportation Research Part A 40 (2006) 572586 www.elsevier.

com/locate/tra

Inter-city bus routing and timetable setting under stochastic demands


Shangyao Yan *, Chin-Jen Chi, Ching-Hui Tang
Department of Civil Engineering, National Central University, Chungli 32054, Taiwan Received 19 April 2004; received in revised form 2 November 2005; accepted 17 November 2005

Abstract Vehicle eet routing and timetable setting are essential to the enhancement of an inter-city bus carriers operating cost, prot, level of service and competitiveness in the market. In past research the average passenger demand has usually served as input in the production of the nal eet routes and timetables, meaning that stochastic disturbances arising from variations in daily passenger demand in actual operations are neglected. To incorporate the stochastic disturbances of daily passenger demands that occur in actual operations, in this research, we established a stochastic-demand scheduling model. We applied a simulation technique, coupled with link-based and path-based routing strategies, to develop two heuristic algorithms to solve the model. To evaluate the performance of the proposed model and the two solution algorithms, we developed an evaluation method. The test results, regarding a major Taiwan inter-city bus operation, were good, showing that the model and the solution algorithms could be useful in practice. 2005 Elsevier Ltd. All rights reserved.
Keywords: Inter-city bus; Fleet routing, Timetable; Stochastic demand; Simulation

1. Introduction Timetable setting and bus routing/scheduling have been the fundamental focus of inter-city bus carriers, because they are essential to a carriers protability, its level of service and its competitiveness in the market. Many factors have been considered in past research into inter-city bus routing and scheduling. These factors generally include the passenger trip demand, the passenger ticket price, the operating constraints, the operating costs, as well as bus maintenance and crew scheduling (Yan and Chen, 2002). Past research usually has used xed parameters, including the projected passenger demand, for bus routing/scheduling, however, in actual operations passenger demand usually varies. The projected demand may not reect the actual daily passenger demand where stochastic disturbances may occur. A planned schedule is the basis for the real future operations. Real operations must fulll the planning objectives by implementing the planned schedule. Thus, the interrelationship between the planned schedule
*

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

0965-8564/$ - see front matter 2005 Elsevier Ltd. All rights reserved. doi:10.1016/j.tra.2005.11.006

S. Yan et al. / Transportation Research Part A 40 (2006) 572586

573

and the real operations must be kept in mind when dealing with real problems with stochastic passenger demands. When these real stochastic demands are not considered, then deterministic-demand models, based on the average (or projected) demand, will tend to use resources too tightly, resulting in an overly optimistic optimal schedule. Although this schedule may be shown to be good in the planning stage, it may produce larger variations in performance when applied in real operations, where stochastic demands often occur. In the worst case scenario, where demand uctuates wildly during daily operations, then the planned schedule could be disturbed enough to lose its optimality. Therefore, to set a good bus route/schedule, not only does the eet and related supply have to be considered, but stochastic passenger uctuations in actual operations also have to be taken into account. Currently, most inter-city bus carriers in Taiwan utilize the trial-and-error process for bus routing and scheduling practices, as follows. Planners adjust a given draft timetable and the bus routes/schedules, considering the number of available buses, their average operating speed, the turn-around time at the station, the bus balance at each station, and the related cost/revenue of bus movements between cities. The draft timetable is typically designed based upon experience, in accordance with projected future passenger demand, the projected market share, and the given right of way. After adjustments, the schedule is then checked to t bus maintenance and crew scheduling need, with possible minor revisions. This process is iterated manually until a desirable timetable and eet routes are obtained. Since the current bus scheduling and routing of Taiwan inter-city bus carriers is neither ecient nor eective, Yan and Chen (2002) recently developed a deterministic scheduling model, with the objective of maxizing the system prot, given a xed projected passenger demand and the operating constraints. Their model was shown to be more systematic and ecient than the traditional trial-and-error method. However, it was established based on the average passenger demand as input. The nal timetable and eet routes that were produced neglected stochastic variations in daily passenger demand that occur in actual operations. Other than Yan and Chen (2002)s work, not much past research on the improvement of inter-city bus scheduling has been found. Salzborn (1980) introduced some rules for scheduling for a bus system for an inter-town route linking a string of interchanges, each of which is the center of a set of feeder routes. Berkhout (1985) provided a structure method which integrated both manual and automatic processes to provide a timetable. However, similar to Yan and Chen (2002), their models assumed the passenger demand to be xed, and neglected the variation of passenger demand in actual daily operations. As a result, their timetables and bus routes/schedules may not be eective for stochastic passenger demands, possibly being inaccurate and inecient in actual operations. It should be mentioned that aside from inter-city bus routing/scheduling, there has been comparatively more research on urban bus routing/scheduling, which can be categorized as bus routing, scheduling, or a combination of routing and scheduling problems. For examples of recent research, see Han and Wilson (1982), Kocur and Hendrickson (1982), Tsao and Schonfeld (1983, 1984), Chua (1984), Koutsopoulos et al. (1985), Ceder and Wilson (1986), Kuah and Perl (1988), van Nes et al. (1988), Chang (1990), Chang and Schonfeld (1991), Ceder (1991), Spasovic et al. (1994), Sinclair and van Oudheusden (1997), Pattnaik et al. (1998), and Prakash et al. (1999). However, their problem characteristics are not the same as for inter-city bus routing/schedluing. As a result, their models and solution solutions are not applicable to our problem. To the best of the authors knowledge, there has been no research on inter-city bus routing and scheduling under stochastic demands, although planning problems with stochastic disturbances have been studied in other elds. For example, Mulvey and Ruszczynski (1995) developed a scenario decomposition method for large stochastic problems. They used two examples, a nancial planning problem and an air freight scheduling problem, to demonstrate the performance of their method. Du and Hall (1997) considered the eet sizing and empty equipment redistribution problem for a center-terminal network with stochastic demands. Yan et al. (2002) analyzed the interrelationship between static and real-time gate assignments, as aected by the stochastic ight delays that occur in real operations. However, they did not further solve for the optimally planned gate assignments subject to stochastic ight arrival/departure disturbances. Kenyon and Morton (2003) considered a stochastic vehicle routing problem (SVRP) with random travel and service times. They provided two models, with the dierent objectives of minimizing the expected completion time and of maximizing the probability that the operation is complete on or before a prespecied target time. List et al. (2003) discussed the eet sizing problem under the uncertainty of future demands. Their formulation focused on robust optimization, using a partial moment measure of risk.

574

S. Yan et al. / Transportation Research Part A 40 (2006) 572586

In this research stochastic disturbances of daily passenger demands that occur in actual operations are considered from the basis of the carriers perspective. A stochastic-demand scheduling model (SDSM) was established by modifying the deterministic-demand scheduling model (DDSM) developed by Yan and Chen (2002). We also applied a simulation technique, coupled with link-based and path-based routing strategies, to develop two heuristic algorithms for solving the SDSM. To compare the performance of the two models and the two solution algorithms, under the stochastic demands that occur in actual operations, we developed a simulationbased evaluation method. In practice, aside from passenger demand, stochastic eet delays may occur in actual operations. One major Taiwan inter-city bus carrier adds a buer time between two connected bus trips to absorb minor delays (Yan and Chen, 2002). For simplicity, stochastic eet delays are not considered here. In addition, we only address the disturbance due to stochastic demands, rather than other types of larger disturbances which cannot be absorbed by a buer time, such as a temporary shortage of vehicles, vehicle malfunctions, unexpected trac incidents or congestion. The incorporation of such disturbances into the model could be a direction of future research. Due to the fact that weekend carrier services are similar to weekday services, as noted in Yan and Chen (2002), we set planning period to be one day, but this is adjustable. Although, in practice, the scheduling process is related to bus maintenance and the crew scheduling processes, these processes are carried out after bus routing and scheduling has been completed. According to the studied Taiwan inter-city bus carrier, in practice the maintenance and crew constraints are rather exible, due to the use of stand-by crews and spare buses. In only a few cases do routes and schedules may have to be slightly modied to meet maintenance or crew scheduling issues. To facilitate problem solving, we therefore exclude these constraints in our modeling. The rest of this paper is organized as follows: In Section 2, the SDSM is introduced. In Section 3, the solution and evaluation methods are developed. In Section 4, numerical tests are performed to evaluate the performance of the models and the solution algorithms. Finally, discussions and conclusions are oered in Section 5. 2. The SDSM Yan and Chen (2002) employed network ow techniques to construct the DDSM. The two major elements in the DDSM, the eet-ow and the passenger-ow networks, respectively, indicate the eet and passenger movements within a specied time period and space locations. There are three types of arcs in the eet-ow network: (1) service arcs: representing a bus trip from one station to another; (2) holding arcs: representing the holding of buses at a station in a time window; and (3) cycle arcs: showing the continuity between two consecutive planning periods. Each passenger-ow network represents a specic OD pair selected from the origin-destination table (known as the OD table). As well, there are three types of arcs in the passenger-ow networks: (1) delivery arcs: representing the transport of passengers from one station to another on a bus trip; (2) holding arcs: indicating the stay of passengers at a station in a given time window; (3) demand arcs: denoting the actual demand for this OD pair to be served in the network. In particular, the demand arc ows upper bound in each passenger-ow network represents the demand for the corresponding OD pair for a given time interval (i.e., 1 h in this research). To respond to daily stochastic disturbances due to uctuations in passenger demand in actual operations, we modify every demand arc ows upper bound in the SDSM to be a random variable. Please refer to Yan and Chen (2002) for a more detailed description of the DDSM. The two-stage stochastic programming concept has been applied to the modeling of stochastic problems, for example, see Higle and Sen (1996) and Birge and Louveaux (1997). This concept is adopted to develop the SDSM in this research. In a general two-stage stochastic optimization process, the decision variables are divided into two groups. In the rst stage, a number of decisions have to be determined before the realization of random variables becomes known. In the second stage, a number of decisions can be determined after the values of the random variables are realized. Similar to the Mulvey and Ruszczynski (1995) air freight scheduling model, we select S stochastic passenger demand scenarios, where S can represent all stochastic events and is subject to testing. That is, we have the S ow upper bounds for every demand arc in each passenger-ow network. In this research, we set the rst-stage decision variables to be the eet-ow variables, Xijs, which have to be determined before the daily passenger demands in actual operations can be known. The sec;d ond-stage decision variables are the passenger-ow variables, Y s ij s, which are used to formulate passenger ;d movements to meet the actual demands after they become known. Note that Y s ij is dened as the arc (i, j) ow

S. Yan et al. / Transportation Research Part A 40 (2006) 572586

575

in the dth passenger-ow network for the sth scenario. Before introducing the model formulation, we list the notation and symbols used in the model formulation: Cij s R(X, s) Ps F Uij Td ij d Wij
;d Hs ij

Ud ij N,A SE CF D Qd,Bd SA BSd BDd

the arc (i, j) cost in the eet-ow network; the sth scenario; the total passenger ow cost for the sth scenario under the eet ows obtained in the rst stage; the probability for the sth scenario; the number of available buses in the eet-ow network; the arc (i, j) ows upper bound in the eet-ow network (note that Uij equals 1 for the service arcs and is a positive integer value for the holding and cycle arcs); the arc (i, j) cost (the negative ticket fare) in the dth passenger-ow network; the dth passenger-ow network (i.e, the dth OD pair); the bus capacity for a service arc (i j) in the eet-ow network (note that a planned load factor could be used in the planning stage); the ows upper bound for the demand arc (i, j) in the dth passenger-ow network, for the sth scenario; the ows upper bound for the holding arc or delivery arc (i, j) in the dth passenger-ow network (note that U d ij is unrelated to the sth scenario; for the holding arc it is the stations passenger capacity and for the delivery arc it is the bus capacity); the set of all nodes and arcs in the eet-ow network; the set of all scenarios; the set of all cycle arcs in the eet-ow network; the set of all passenger-ow network (i.e., all ODs); the set of all nodes and arcs in the dth passenger-ow network; the set of all service arcs in the eet-ow network, which is the same as the set of all delivery arcs in all passenger-ow networks; the set of all demand arcs in the dth passenger-ow network; the set of holding arcs and delivery arcs in the dth passenger-ow network.

Note that Bd = BSd [ BDd. The formulation of the SDSM is then: X X Min C ij X ij P s RX ; s s.t. X
j 2N ij2A

1 2 3

X ij

X
k 2N

s2SE

X ki 0;

8i 2 N ;

X ij 6 F ; 8ij 2 A; 8ij 2 A;

ij2CF

0 6 X ij 6 U ij ; X ij 2 Integer; where RX ; s Min

4 5

X X X
j2Qd d 2D ij2Bd

s; d Td ij Y ij

6
;d Ys ki 0;

;d Ys ij

X
k 2Q d

8i 2 Qd ; 8d 2 D;

7 8 9 10

X
d 2D

;d Ys ij 6 W ij X ij ;

8ij 2 SA; 8ij 2 BS d ; 8d 2 D;

;d s; d 0 6 Ys ij 6 H ij ;

06

;d Ys ij

Ud ij ;

8ij 2 BD ; 8d 2 D.

576

S. Yan et al. / Transportation Research Part A 40 (2006) 572586

The objective function (1) denotes the cost minimization of the eet ows plus the expected cost (including a negative revenue for passenger fares) of the passenger ows taken over all scenarios. Since the ticket fare from the passenger-ow networks is included in the form of a negative cost, this objective is equivalent to a maximization of the prot. In the rst stage, constraint (2) is the ow conservation constraint at every node in the eet-ow network. Constraint (3) denotes that the number of buses used in the eet-ow network should not exceed the available number of buses. Constraint (4) denotes that all arc ows in the eet-ow network are kept within their bounds. Constraint (5) ensures that all eet ows are integers. In the second stage, the objective function (6) represents the cost minimization of the passenger ows for the sth scenario under the eet ows obtained in the rst stage. Constraint (7) is the ow conservation constraint at every node in each passenger-ow network, for the sth scenario. Constraint (8) ensures that the passenger delivery volume for the sth scenario is kept within the buss capacity. Constraint (9) holds that all demand arc ows in each passengerow network are kept within their bounds, for the sth scenario. Constraint (10) ensures that for the sth scenario all holding and delivery arc ows in each passenger-ow network are kept within their bounds. Note that, as in Ruszczynski and Shapiro (2003), an alternative formulation of the SDSM can be made by 2 s splitting the rst stage decision variable Xijs into X 1 ij s; X ij s; . . . ; X ij s corresponding to scenarios, and then adding the non-anticipativity constraints into the model. Note that X s ij is dened as the arc (i, j) ow in the eetow network for the sth scenario. Thus, we can reformulate the SDSM as follows: 0 1 X X X X s; d A 11 P s@ C ij X s Td Min ij ij Y ij s.t. X X
j 2Q
d

s2SE

ij2A

Xs ij

X
k 2N

d 2D ij2Bd

Xs ki 0;
;d Ys ki 0;

8i 2 N ; s 2 SE; 8i 2 Qd ; 8d 2 D; s 2 SE;

12 13 14 15 16 17 18 19 20 Xs ij s

j 2N

;d Ys ij

X
d

k 2Q

Xs ij 6 F ; s 2 SE ;
;d s Ys ij 6 W ij X ij ;

X
d 2D

ij2CF

8ij 2 SA; s 2 SE;

0 6 Xs ij 6 U ij ; 06 06 Xs ij
;d Ys ij ;d Ys ij

8ij 2 A; s 2 SE; 8ij 2 BS ; 8ij 2 BD ;


d d

6 6

;d Hs ij ;

8d 2 D; s 2 SE; 8d 2 D; s 2 SE;

Ud ij ;

2 Integer;

8ij 2 A; s 2 SE;

t Xs ij X ij ;

8ij 2 A; s; t 2 SE; s 6 t.

The non-anticipativity constraints (20) indicate that the rst stage variables have to be equal in each scenario. It should be mentioned that if we relax the non-anticipativity constraints in the SDSM, then the optimal solutions, for all scenarios, can be respectively solved by the DDSM. As indicated in Birge and Louveaux (1997) and Ruszczynski and Shapiro (2003), the average of the scenarios optimal objective values (i.e., the expected value) is less than or equal to the optimal SDSM value. That is, the average of all scenarios objective values provides the lower bound of the optimal SDSM solution. 3. Solution algorithms and evaluation method Two solution algorithms, based on two routing strategies and a simulation technique, are developed to solve the SDSM. The solution processes of the two algorithms are the same. We rst relax the non-anticipativity constraints (20) to generate S independent DDSMs. Then, using the mathematical programming solver, CPLEX, we solve S DDSMs for bus routs/schedules. Note that each DDSM corresponds to a passenger demand scenario obtained from the simulation, with a demand distribution similar to that in the real operations. As mentioned previously, the average objective value of the S scenarios is the lower bound of the opti-

S. Yan et al. / Transportation Research Part A 40 (2006) 572586

577

mal solution. Thereafter, based on link-based and path-based routing strategies, we develop two algorithms for the selection of the bus trips/routes which could best absorb stochastic demand disturbances in the S scenarios. Note that a path (i.e., a sequence of stations) presented in the path-based routing strategy denotes a bus route. In the next iteration, by xing the selected bus trips/routes (which is a part of bus eet ows), we again resolve the DDSM the same way. The process is not stopped until the number of iterations exceeds a preset number. Finally, we x all selected bus trips/routes obtained previously and apply the average demand of the S scenarios as input to solve the DDSM for the rest of bus trips/routes. It should be mentioned that the two algorithms produce near-optimal solutions. This primarily helps the carriers to formulate a better bus route/schedule, which can eectively meet stochastic passenger demands. The stopping rule for selecting bus trips/routes is adjustable. For example, the change rate of the selected bus trips/routes in the solution process could be used as the stopping criterion. The two solution algorithms are outlined as follows. 3.1. Link-based heuristic algorithm (LBHA) As introduced above, we have to select the bus trips that could best absorb stochastic demand disturbances. To do this, if the number of times a bus trip is served during S scenarios exceeds a required appearance rate (RAR), then we will select this bus trip. For example, suppose that two bus trips, A and B, appear in 20 and 10 scenarios respectively, in 30 simulation scenarios. If we assume that the RAR is set to be 50%, then bus trip As rate (20/30 = 66%) is larger than the RAR, so we select bus trip A as part at the nal solution. On the other hand, bus trip Bs rate (10/30 = 33%) is smaller than the RAR, so we do not select bus trip B. Note that the determination of the RAR is subject to testing. The steps of the LBHA are: Set S (a preset number of scenarios), M (a preset number of iterations), and the RAR. Randomly generate S passenger demand scenarios. Set iteration m = 1. Set scenario s = 1. Solve the DDSM with the passenger demand of the sth scenario to obtain the eet ows. If s > S, then go to Step 6. Otherwise, s = s + 1 and return to Step 4. Calculate the number of times each bus trip is served. If the value of the appearance rate is larger than the RAR, then select this bus trip. Step 7: If m > M, then go to Step 8. Otherwise, x the selected bus trips obtained from previous iterations, m = m + 1, and return to Step 3. Step 8: Fix all selected bus trips obtained from previous iterations and apply the average demand of S scenarios as input to solve for the rest of the bus trips. To do this, we can rst x the selected service arc variables and then solve the rest of the DDSM. Step 9: Use the ow decomposition algorithm (Yan and Lin, 1997) to decompose the link ows into arc chains, each denoting a buss daily route/schedule. Step Step Step Step Step Step Step 0: 1: 2: 3: 4: 5: 6:

3.2. Path-based heuristic algorithms (PBHA) In each scenario, after solving the DDSM to obtain the eet ows, unlike the LBHA, we now use the ow decomposition algorithm (Yan and Lin, 1997) to obtain a route for each bus. We select bus routes that could absorb the stochastic demand disturbances. As described above, the path ow pattern obtained by using the ow decomposition method may not be unique, which makes it dicult or unsuitable to nd a bus route for which the appearance rate in S scenarios is larger than the RAR. Therefore, we use a denition of path similarity, as well as a selection strategy, to select the bus routes. The determination of similar paths is based on the sequence of stations along a bus route. An example of path similarity is illustrated as follows. As shown in Fig. 1, assume that there are four paths (A, B, C, and D), and that path A is the basis for comparison, meaning that paths B, C, and D are compared with path A for

578

S. Yan et al. / Transportation Research Part A 40 (2006) 572586

basic path

a similar path

not a similar path

not a similar path

Fig. 1. Demonstration of path similarity.

similarities. For example taking path B, the stations along route (2 ! 1 ! 5 ! 3 ! 2 ! 4) contain all path A stations (1 ! 3 ! 2 ! 4). In addition, the sequence of path As stations is the same as a part of path B. Thus, path B is dened as similar to path A. Taking another example, although the stations of path C (1 ! 2 ! 3 ! 4) are included in path A, the sequence of path Cs stations is not the same as in path A, so path C is not considered to be similar to path A. As for path D (2 ! 1 ! 3 ! 2 ! 5), because it does not include all stations in path A (station 4 is lacking), it is not similar to path A. This denition of path similarity is suitable for the Taiwan inter-city bus network, which is a linear network. The users may set their own criteria, based on the network type and other considerations. Using this path similarity denition, we now develop a strategy for path selection which can absorb stochastic demand disturbances. We rst calculate for each scenario the average number of stations that a bus serves, then select the basic scenario, the scenario with the minimum average number of stations. Secondly, we set all the bus paths in the basic scenario to be basic paths. The basic scenario and the basic paths are primarily used for comparison with the bus paths in the other scenarios. Note that the setting of the basic scenario and of the basic paths is adjustable. Next, we compare each basic path with the bus paths in the other S 1 scenarios, to nd the similar paths. In each scenario, if a similar path to the basic path is found, then we stop comparing the basic path with the other bus paths in the scenario and then move on to compare another scenario. That is, a basic path has at most S 1 similar paths during S 1 scenarios. Finally, if a basic path has S 1 simialr paths, then we select a representative path from S (including the basic path) similar paths. A representative path from S similar paths is designated as a bus route that can potentially absorb the stochastic demand disturbances. Because a representative path will be xed in later iterations of the solution process, it is not allowed to contain the same service arcs as those in the paths obtained from the pervious iterations. Otherwise, this would violate the upper bound of the service arc ow (i.e., one) in the eet-ow network when the DDSM is solved in the next iteration. Therefore, for each similar path we calculate the dierence between the number of service arcs and the average number of service arcs in S similar paths. We then select the most similar path with the minimum dierence that conforms to the above restrictions to be the representative path. The steps of the PBHA are: Step Step Step Step Step Step 0: 1: 2: 3: 4: 5: Set S (a preset number of scenarios) and M (a preset number of iterations). Randomly generate S passenger demand scenarios. Set iteration m = 1. Set scenario s = 1. Solve the DDSM with the passenger demand of the sth scenario to obtain the eet ows. Use the ow decomposition algorithm (Yan and Lin, 1997) to decompose the link ows into arc chains, each denoting a buss daily route/schedule.

S. Yan et al. / Transportation Research Part A 40 (2006) 572586

579

Step 6: If s > S, then go to Step 7; otherwise, s = s + 1, and return to Step 4. Step 7: Apply the aforementioned selection strategy to select the representative path. Step 8: If m > M, then go to Step 9; otherwise, x the service arcs of the representative path obtained from the previous iterations, m = m + 1, and return to Step 3. Step 9: Fix the service arcs of all the selected representative paths obtained previously and apply the average demand of the S scenarios as input to solve for the rest of the eet ows. To do this, we rst x the service arc variables that correspond to the selected eet ows and then solve the rest of the DDSM. Step 10: Use a ow decomposition algorithm (Yan and Lin, 1997) to decompose the link ows into arc chains, each denoting a buss daily route/schedule.

3.3. The evaluation method Theoretically, the S scenarios in the SDSM are one possible set of realizations of all stochastic events. The planned S scenarios may dier from the events that actually occur. Therefore, we use a simulation technique to develop an evaluation method. In this evaluation method, we rst randomly generate D passenger demand evaluation days for each OD pair. Note that the passenger demand data for D evaluation days are dierent from that for the S scenarios used in Sections 3.1 and 3.2. Then, using the timetables obtained from the DDSM and the SDSM, we respectively solve for the optimal passenger ows and the system prot, for each evaluation day. For ease of writing, we refer to the planned S scenarios as the planning stage and the D evaluation days as the evaluation stage. Finally, we perform a comparison of the planning and the evaluation stages of the DDSM and the SDSM. The steps of the evaluation method are listed as follows: Step Step Step Step Set D (a preset number of evaluation days). Randomly generate D passenger demand evaluation days. Set evaluation day d = 1. Use the timetables obtained from the DDSM and the SDSM with the two heuristic algorithms, to respectively solve for the optimal passenger ows, to obtain the system prot for the dth evaluation day. To do this, we rst x the eet-ow network variables that correspond to the obtained eet ows and then solve the rest of the DDSM. Step 4: If d > D, then go to Step 5; otherwise, d = d + 1, and return to Step 3. Step 5: Compare the variations between the planning and the evaluation stages. Calculate the statistical results for D evaluation days, including the average objective value and the average percentage of demand served. 0: 1: 2: 3:

4. Numerical tests To test how well the model and the solution algorithms may be applied in the real world, we performed some numerical tests. To understand the inuence of the stochastic demand disturbances on the solution, we also performed sensitivity analyses on the mean and the standard deviation of the passenger demand. We used the C computer language to write the necessary programs, coupled with the CPLEX 7.1 mathematical programming solver, to solve the problem. The tests were performed on an Intel P4 2G with 1GB RAM in the environment of Microsoft Windows 2000. 4.1. Data analysis and test results Our numerical tests were mainly based on data from a major Taiwan inter-city bus carrier, from October of 1998. There were ve cities (Taipei, Taichung, Chiayi, Tainan, and Kaohsiung) connected along a freeway, to form a linear network. Note that the model and the solution algorithms developed in this research can be applied to other types of networks, with suitable modications. The eet size and the planned load factor

580 Table 1 LBHA and PBHA test results Case

S. Yan et al. / Transportation Research Part A 40 (2006) 572586

Lower bound of optimal solution (NT$)

Objective value (NT$) LBHA 50% RAR 60% RAR 2333754 2326186 2313941 2296328 2290232 2286702 2309032 2298822 2309591 2299432 PBHA

WEG (%) LBHA 50% RAR 0.75 0.95 0.04 0.16 0.25 60% RAR 0.36 1.11 1.41 2.36 2.30 2.37 1.84 2.06 1.79 1.91 PBHA

1 2 3 4 5

2342283 2352208 2347145 2351793 2344090

2324766 2329747 2346197 2348001 2338191

for each bus were set to be 170 and 0.6, respectively, in accordance with the carrier operations. Several types of buses, with similar capacities (3337 seats, or an average of 35 seats) composed the eet. For simplicity, we assume the buses to be the same size, 35 seats per bus. The distribution of the passenger demand, including the mean and the standard deviation of the passenger demand for each OD pair, followed a truncated normal distribution. Specically, the total average demand of all OD pairs was about 11,033 trips per day. For example, for the TaipeiTaichung pair, the average demand and the standard deviation were about 2556 and 51 trips per day, respectively. Note that the demand distribution is adjustable for dierent market conditions. All the cost parameters and other input, such as the right of way, the travel distance, the travel time, the turn-around time, the speed, and the ticket price, were primarily based on actual operating data, with reasonable simplications. Before starting to perform numerical tests, we evaluated a suitable number of scenarios and evaluation days (i.e., S and D). We tested dierent numbers of scenarios and evaluation days. We found that, after 30 scenarios or evaluation days the bus route/schedule results were similar, but the computational times increased rapidly. For ease of testing, the number of scenarios and evaluation days were both set to be 30, but these are adjustable. Carriers can use dierent numbers of scenarios and days for their own tests. Five dierent cases of passenger demands, each with 30 scenarios and 30 evaluation days, were tested to compare the LBHA, PBHA, and, DDSM. Each case used the average passenger demand, for the 30 scenarios, as input to solve the DDSM. We chose two RARs, 50% and 60%, for the LBHA. In addition, we also evaluated a suitable number of iterations (i.e., M) during the solution process. The test results indicated that 1 and 3 interations for the LBHA and the PBHA were the best, respectively. To measure the eectiveness of the heuristic algorithms, we estimated the worst error gap (WEG) of the approximate solution. The WEG was dened as follows: Objective value of LBHA=PBHA Lower bound of optimal solution . WEG% 21 Lower bound of optimal solution As shown in Table 1, we found that the LBHA with the 50% RAR, yielded the best near-optimal solutions for most of the ve cases, while the PBHA generally oered the worst ones. The average WEGs of the ve LBHA cases, with the two RARs, and the PBHA were about 0.43%, 1.51%, and 1.99%, respectively. We also found that the WEGs for the LBHA and the PBHA were all within 2.37%, for all ve cases. Especially in Case 3, for the LBHA with the 50% RAR, the WEG was only 0.04%. These results indicate that the heuristic solutions obtained in this research, applying the LBHA and the PBHA, were very close to the optimal SDSM solution, which shows the good performance of our algorithms for the solving of stochastic-demand scheduling problems. The computation times for the LBHA with the 50% and 60% RAR were within 90 min, for all ve cases. Although the PBHA was more time-consuming to solve than the LBHA, the PBHA was still solvable within 200 min, for all ve cases. These results indicate that the solution algorithms developed in this research could be applied in actual operations. 4.2. Evaluation results Using the evaluation method, we compared the DDSMs performance after 30 evaluation days with the SDSMs. We analyzed the results in terms of two aspects, (1) the objective value and (2) the timetable, which are discussed below.

S. Yan et al. / Transportation Research Part A 40 (2006) 572586

581

4.2.1. Objective value The symbols used are dened as follows: ZS, ZD: the objective value of the SDSM (solved by LBHA or PBHA) and the DDSM for the planning stage; E ZE S ; Z D : the objective value of the SDSM (solved by LBHA or PBHA) and the DDSM for the evaluation stage. Then, V (%), G (%), and I (%) are dened as follows: E Z S or Z D Z E S or Z D V % E ; ZE S or Z D ZS ZD G% ; ZD ZE ZE I % S E D . ZD

22 23 24

E V, G, and I indicate the variation between the ZS (or ZD) and Z E S (or Z D ), between the ZS and ZD, and between E E E the Z S and Z D , respectively. A positive G (or I) denotes that the ZS (or Z E S ) is better than the ZD (or Z D ), and vice verse. As shown in Table 2, we found that most Vs for the LBHA with the 50% and 60% RAR and the PBHA were less than those for the DDSM, meaning that the obtained DDSM plan will have a larger variation between the planning and the evaluation stages. The average Vs for the LBHA with the 50% and 60% RAR, the PBHA, and the DDSM, were 0.39%, 1.31%, 0.83%, and 1.52%, respectively. Table 3 shows that for all ve cases, the Gs for the LBHA with the 50% and 60% RAR, and the PBHA, were all negative, meaning that, in terms of planning results the SDSM is worse than DDSM. Here, the reader should not be misled to assume that the DDSM is better than the SDSM, although the DDSM does yield a better planning objective

Table 2 V (%) results Case SDSM LBHA with 50% RAR 1 2 3 4 5 Average 0.43 0.50 0.76 0.20 0.04 0.39 LBHA with 60% RAR 0.40 0.53 1.39 1.70 2.53 1.31 PBHA 0.82 0.60 1.17 0.50 1.08 0.83 1.52 1.97 0.98 1.83 1.33 1.52 DDSM

Table 3 G (%) and I (%) results Case G (%) LBHA with 50% RAR 1 2 3 4 5 Average 1.71 1.10 0.07 0.38 0.19 0.69 LBHA with 60% RAR 1.33 1.25 1.44 2.58 2.23 1.77 PBHA 3.32 1.98 2.08 2.01 1.84 2.25 I (%) LBHA with 50% RAR 0.22 0.34 1.69 1.64 1.09 0.91 LBHA with 60% RAR 0.23 1.22 0.93 0.92 1.63 0.89 PBHA 1.04 0.54 0.05 0.28 0.54 0.08

582

S. Yan et al. / Transportation Research Part A 40 (2006) 572586

value. The actual performance of the planning models should only be evaluated after the results are applied in real operations. As reasoned previously, the Is were all positive, except for Case 1, meaning that the optimal DDSM plan will lose its optimality and become inferior to the SDSMs, when applied to actual operations where stochastic demands occur. 4.2.2. Timetable We further compared the dierence between the SDSM and the DDSM timetables. Case 3, the LBHA with the 50% RAR had the best I (1.69%), so it was used for the comparison. Two OD pairs, the Taipei Kaohsiung pair (a high passenger demand OD pair) and the TaipeiChiayi pair (a low passenger demand OD pair), were discussed. Note that high and low passenger demands are relative and are only for comparison purposes. As shown in Table 4, we found that for the TaipeiKaohsiung pair, the number of bus trips suggested by the SDSM (41) was greater than for the DDSM (39). As well, the percentage of demand served was higher for the SDSM than for the DDSM, the average increase being 1.91% after 30 evaluation days. During peak hours, most SDSM and DDSM departure intervals were similar, about 15 min, however, during non-peak hours, the SDSM departure intervals were mostly less than those for the DDSM, being 4560 min rather than 4575 min. Similar results were also found for the other high passenger demand OD pairs, meaning both the SDSM and the DDSM set a 15-min departure interval for peak hours; but for non-peak hours the SDSM yielded a shorter and more suitable departure interval (e.g., 4560 min for the TaipeiKaohsiung pair) than did the DDSM. The number of bus trips suggested by the SDSM for the TaipeiChiayi pair (22) was less than that for the DDSM (24). However, the percentage of demand served by the SDSM was still higher than for the DDSM, the average increase being 1.27%, after 30 evaluation days. This is because, given the average passenger demand, the DDSM incorporated a set of inappropriate departure intervals in its timetable, which means it does not eectively serve the stochastic demands. On the other hand, the SDSM yielded an eective timetable, with a set of departure intervals that appropriately matched the stochastic demands. For example, the DDSM scheduled ve bus trips before 9:00 a.m., at 6:00 a.m., 7:00 a.m., 7:15 a.m., 8:15 a.m., and 9:00 a.m., while the SDSM only had four, at 6:45 a.m., 7:00 a.m., 8:00 a.m., and 9:00 a.m. However, in this time period, the percentage of demand served by the SDSM was higher than that met the DDSM, the average increase being 0.98% after 30 evaluation days. In addition, for the TaipeiChiayi pair, the SDSM yielded a more

Table 4 Timetables for TaipeiKaohsiung and TaipeiChiayi pairs TaipeiKaohsiung SDSM (solved by LBHA with 50% RAR) 06:00a 06:30 07:00 07:15 07:30 07:45 08:00 08:15 09:00 09:30 09:45 10:00 10:15 10:30
a b

TaipeiChiayi DDSM SDSM (solved by LBHA with 50% RAR) 11:30 12:15 13:15 13:30 14:00 15:15 16:00 16:15 16:30 17:00 17:45 18:00 18:15 19:15 19:30 20:00 20:45 21:00 21:45 22:00 22:45 23:00 23:30 00:00b 06:00a 07:00 08:00 09:00 10:00 11:00 12:30 13:45 15:00 16:00 17:00 17:30 18:00 19:00 20:00 21:00 22:00 23:00 23:45 00:00 01:00 01:30b DDSM

11:00 12:00 12:15 12:30 13:00 13:15 14:00 15:00 15:15 16:00 16:15 17:00 17:15 17:30

18:00 18:15 19:00 19:15 19:45 20:00 21:00 21:45 22:00 22:45 23:00 23:45 00:00b

06:30a 06:45 07:15 07:30 07:45 08:00 08:15 08:30 09:15 10:00 10:15 10:30 11:00 11:15

06:00a 07:00 07:15 08:15 09:00 09:45 11:00 11:30 12:30 14:00 15:00 16:15 17:00 18:30

18:45 19:15 20:00 22:00 22:15 23:30 23:45 00:00 01:00 01:30b

The rst bus trip. The last bus trip.

S. Yan et al. / Transportation Research Part A 40 (2006) 572586

583

uniform timetable than did the DDSM. Most of the departure intervals for the SDSM were 60 min, irregardless of whether for peak or non-peak hours, but the DDSM departure intervals varied from 15 to 90 min. Similar results were also found for the other low passenger demand OD pairs, meaning that a timetable with a uniform departure interval is more suitable for a low passenger demand OD pair (e.g., 60 min for the Taipei Chiayi pair). 4.3. Mean passenger demand We tested two more situations, 50% and 150% of the original mean. Sensitivity analyses of the other situations can be performed in a similar manner. We analyzed the results in terms of two aspects: (1) the I results and (2) the timetable, which are discussed below. 4.3.1. I results As shown in Table 5, most Is were positive for all three situations (including the original mean as shown in Table 3), indicating that the SDSM could be superior to the DDSM when the mean is dierent. However, the superiority associated with a change of the mean showed no specic trend. For example, in the 50% mean situation, only the LBHA with the 60% RAR had positive Is for all ve cases. In the original mean situation, the LBHA with the 50% and 60% RAR, and the PBHA, had positive Is for all ve cases, except for Case 1. When the mean was 150%, the LBHA with the 60% RAR and the PBHA had positive Is for all ve cases, except for Cases 5 and 1, respectively. 4.3.2. Timetable In the 50% mean situation, Case 1, the SDSM solved by the LBHA with the 60% RAR had the best I (1.58%), so it was used for the comparison. Similarly, when the mean was 150%, Case 2 was used to compare the SDSM (solved by the PBHA) and the DDSM results. For a 50% mean, we found that the SDSM yielded a more uniform timetable for all OD pairs than did the DDSM. For example, most of the SDSM departure intervals for the TaipeiKaohsiung pair were 60 min, while for the DDSM they varied from 30 to 105 min. This result shows that for the 50% mean situation, a timetable with a uniform departure interval (e.g., 60 min for the TaipeiKaohsiung pair) is more suitable, which is similar to that for low passenger demand OD pairs (discussed in Section 4.2) under the original mean. With 150% mean, there was no specic dierence trend between the SDSM and the DDSM departure intervals. We only found that for each high passenger demand OD pair, the number of SDSM bus trips was signicantly greater than for the DDSM, especially during peak hours. However, for each low passenger demand OD pair, the number of SDSM bus trips was slightly greater than or equal to the DDSM ones. This is most likely because when all demands increase to 150% mean, then more bus trips for a high demand OD pair is more benecial than for a low demand one, in peak hours, given the available buses constraint in short term operations.

Table 5 I (%) results for dierent means Case 1 50% Mean LBHA with 50% RAR LBHA with 60% RAR PBHA 150% Mean LBHA with 50% RAR LBHA with 60% RAR PBHA 0.37 1.58 2.58 1.76 0.02 0.05 Case 2 0.24 0.13 0.23 0.29 1.57 1.79 Case 3 0.27 0.12 0.88 0.84 0.96 0.79 Case 4 0.64 0.13 1.89 1.36 0.25 0.83 Case 5 0.78 1.57 0.61 1.62 0.04 0.49 Average 0.04 0.71 1.00 1.18 0.55 0.77

584

S. Yan et al. / Transportation Research Part A 40 (2006) 572586

Table 6 I (%) results for dierent standard deviations (LBHA) Case 1 50% Standard deviation 50% RAR 0.69 60% RAR 0.86 150% Standard deviation 50% RAR 1.21 60% RAR 1.18 Case 2 0.28 0.25 1.24 0.89 Case 3 0.13 0.09 1.55 1.08 Case 4 0.11 0.24 1.09 1.05 Case 5 0.54 0.30 0.72 0.35 Average 0.24 0.25 0.87 0.77

4.4. Standard deviation We also tested two more situations, 50% and 150% of the original standard deviation and analyzed the results in terms of two aspects: (1) the I results and (2) the timetable, which are discussed below. 4.4.1. I results We found that the SDSM results solved by the LBHA with the 50% or 60% RAR were better than those solved by the PBHA, showing that this LBHA is more useful than the PBHA for solving the SDSM, irregardless of the standard deviations. To save space, only the I results for the SDSM solved by the LBHA with the 50% and 60% RAR are discussed here. As shown in Table 6, most of the Is for the ve cases were positive, showing that the SDSM outperformed the DDSM, irregardless of the standard deviations. The Is for the 50% standard deviation were the smallest. Specically, the average I for the SDSM solved by the LBHA with the 50% or 60% RAR was smaller than 0.25%. This implies that if the demand uctuates slightly during daily operations, then the DDSM might still be useful for schedule planning, resulting in only a slightly worse result than the SDSM. However, for the original and the high demand uctuations, the DDSM yields a result that does not reect the stochastic variations in daily passenger demand, and this disturbance reduces its planned performance in actual operations. 4.4.2. Timetable In the 50% standard deviation, Case 1, the LBHA solved by the 60% RAR had the best I (0.86%), so it was used for the comparison. Similarly, when the standard deviation was 150%, Case 3 was used to compare the SDSM (solved by the LBHA with 50% RAR) with the DDSM results. Given a 50% standard deviation, we found that the timetables for the SDSM and for the DDSM were similar for all OD pairs. For example, for the TaipeiChiayi pair, the number of SDSM and DDSM bus trips served was the same, 23. The departure intervals were also similar, every 30 or 60 min. This implies that a timetable with a regular departure interval (e.g., 30 or 60 min for the TaipeiChiayi pair) is suitable for low uctuations in passenger demand. For the 150% standard deviation, for all OD pairs, the number of SDSM bus trips served was greater than for the DDSM. We also found that for most OD pairs, the additional bus trips were scheduled between 6:00 a.m. and 8:00 a.m., possibly because in Taiwan the demand variance in this time period is greater than in the other time periods. These results imply that a carrier should oer more bus trips, especially in the time period with a larger demand variance, to successfully meet high uctuations in passenger demand. 5. Discussion and conclusions To consider the stochastic disturbances of varying passenger demand that occur in actual operations, we established a stochastic-demand scheduling model, SDSM, based on a deterministic-demand scheduling model, DDSM. We also employed a simulation technique, coupled with link- and path-based routing strategies, to develop two solution algorithms, the LBHA and the PBHA. To compare the performance of the SDSM and the DDSM under the stochastic demands that occur in actual operations, we further developed a simulation-based evaluation method. Numerical tests, including two sensitivity analyses on the mean and

S. Yan et al. / Transportation Research Part A 40 (2006) 572586

585

the standard deviation of the passenger demand, were performed to evaluate the SDSM and DDSM. The model, the solution algorithm and the test results should serve as a useful reference for transportation planners and operators. Some ndings from the model and solution algorithm development, as well as the test results, are summarized and discussed below: 1. The passenger demand for each time interval for each OD pair in the DDSM is xed (i.e., a projected or an average demand). To respond to daily stochastic passenger demands in actual operations, the xed demand in the SDSM needs to be modied to be a random variable, following a specic distribution to actual operations. Based on two routing strategies and a simulation technique, we developed the LBHA and the PBHA to eciently solve the SDSM, which is more dicult to solve, but better, than the DDSM. Comparing the LBHA and PBHA, we found that dierent demand means or standard deviations could aect the performance. With the current standard deviation but dierent means, there was no specic dierence in their performance. However, for dierent standard deviations, but with the current mean, the LBHA could more eciently solve the SDSM than could the PBHA. Accordingly, when applying the SDSM a carrier should try both the LBHA and PBHA, with dierent parameters, to nd the best option to solve the SDSM, suitable for its own applications. 2. Although the DDSM yields a better objective value than the SDSM in the planning stage, its optimal planned schedule does not incorporate the stochastic passenger demands that often occur in real operations. An optimal DDSM plan will vary greatly in performance, losing its expected optimality when applied in real operations. It is important that the actual performance of dierent planning models should be evaluated after schedules are applied in real operations, instead of only on their planning results. 3. In current Taiwan inter-city market conditions, for a high passenger demand OD pair, both the SDSM and the DDSM yield similar departure intervals for peak hours. For non-peak hours with stochastic disturbances, the SDSM yields a smaller, but more suitable, departure interval than the DDSM. For a low passenger demand OD pair, the SDSM results in a timetable with more uniform departure intervals that is suitable for both peak and non-peak hours. This indicates that for a low passenger demand OD pair, a stochastic model (e.g., the SDSM) is more suitable than a deterministic model (e.g., the DDSM). 4. The demand mean sensitivity analysis indicates that a timetable with a uniform departure interval is more suitable for low demand (e.g., the 50% mean) situations. This is similar to low passenger demand OD pairs given the current demand mean situation, as mentioned in point 3. This implies that it is better to have a uniform departure interval for each low passenger demand OD pair. Under high demand (e.g., the 150% mean), no specic departure interval pattern was found. Moreover, for a high passenger demand OD pair it is found that serving more bus trips is more benecial than for a low demand OD pair, in peak hours, given the available buses constraint in short term operations. This implies that when the demand of all OD pairs increases simultaneously, a carrier could allocate more bus trips to the OD pairs with relatively high demand than other OD pairs. 5. The sensitivity analysis of the standard deviation indicates that if the demand uctuates slightly during daily operations, then the DDSM could be useful enough to plan a schedule, performing only slightly worse than the SDSM. However, for high demand uctuations, the DDSM results will be disturbed enough to reduce their planned performance, making them inferior to the SDSM. This implies that stochastic models are necessary for planning real operations with high demand uctuations. Moreover, in low demand uctuation situations (e.g., the 50% standard deviation), a timetable with a regular departure interval is suitable, but in high demand uctuation situations (e.g., the 150% standard deviation), more bus trips should be served in the time period with a larger demand variance to meet the uctuations in passenger demand. Finally, for dierent demand distributions and dierent service networks, carriers may develop dierent models and explore dierent solution algorithms suitable for their own operations. However, the model and the solution algorithms proposed in this research should be useful for reference. The development of other models and solution algorithms in dierent operating environments could be a direction of future research.

586

S. Yan et al. / Transportation Research Part A 40 (2006) 572586

Acknowledgements This research was supported by a Grant (NSC-92-2211-E-008-047) from the National Science Council of Taiwan. We thank the inter-city bus carrier for providing the test data and their valuable opinions. References
Berkhout, J., 1985. Structure method for vehicle scheduling. Computer Scheduling of Public Transport 2, 199208. Birge, J.R., Louveaux, F., 1997. Introduction to Stochastic Programming. Springer-Verlag, New York. Ceder, A., Wilson, N.H.M., 1986. Bus network design. Transportation Research B 20, 331344. Ceder, A., 1991. Transit scheduling. Journal of Advanced Transportation 25, 137160. Chang, S.K., 1990. Analytic optimization of bus systems in heterogeneous environments. Ph.D. dissertation, University of Maryland, College Park, Maryland, USA. Chang, S.K., Schonfeld, P.M., 1991. Multiple period optimization of bus transit System. Transportation Research B 25, 453478. Chua, T.A., 1984. The planning of urban bus routes and frequencies: a survey. Transportation 12, 147172. Du, Y., Hall, R., 1997. Fleet sizing and empty equipment redistribution for center-terminal transportation networks. Management Science 43, 145157. Han, A.F., Wilson, N.H.M., 1982. The allocation of buses in heavily utilized networks with overlapping routes. Transportation Research B 16, 221232. Higle, J.L., Sen, S., 1996. Stochastic Decomposition. Kluwer Academic publishers, Dordrecht. Kenyon, A.S., Morton, D.P., 2003. Stochastic vehicle routing with random travel times. Transportation Science 37, 6982. Kocur, G., Hendrickson, C., 1982. Design of local bus service with demand equilibrium. Transportation Science 16, 149170. Koutsopoulos, H.N., Odoni, A.R., Wilson, N.H.M., 1985. Determination of headways as a function of time varying characteristics on a transit network. Computer Scheduling of Public Transport 2, 391413. Kuah, G.K., Perl, J., 1988. Optimization of feeder bus routes and bus stop spacing. Journal of Transportation Engineering 114, 341354. List, G.F., Wood, B., Nozick, L.K., Turnquist, M.A., Jones, D.A., Kjeldgaard, E.A., Lawton, C.R., 2003. Robust optimization for eet planning under uncertainty. Transportation Research E 39, 209222. Mulvey, J.M., Ruszczynski, A., 1995. A new scenario decomposition method for large-scale stochastic optimization. Operations Research 43, 477490. Pattnaik, S.B., Mohan, S., Tom, V.M., 1998. Urban bus transit route network design using genetic algorithm. Journal of Transportation Engineering 124, 368375. Prakash, S., Balaji, B.V., Tuteja, D., 1999. Optimizing dead mileage in urban bus routes through a nondominated solution approach. European Journal of Operational Research 114, 465473. Ruszczynski, A., Shapiro, A., 2003. Stochastic Programming. Elsevier, Amsterdam. Sinclair, M., van Oudheusden, D.L., 1997. Network approach to trip frequency scheduling for bus routes in heavily congested cities. European Journal of Operational Research 103, 1827. Spasovic, L.N., Boile, M.P., Bladikas, A.K., 1994. Bus transit service coverage for maximum prot and social welfare. Transportation Research Record 1451, 1222. Salzborn, F.J.M., 1980. Scheduling bus systems with interchanges. Transportation Science 14, 211231. Tsao, S.M., Schonfeld, P.M., 1983. Optimization of zonal transit service. Journal of Transportation Engineering 109, 257272. Tsao, S.M., Schonfeld, P.M., 1984. Branched transit service: an analysis. Journal of Transportation Engineering 110, 112128. van Nes, R., Hamerslag, R., Immer, B.H., 1988. Design of public transport networks. Transportation Research Record 1202, 7483. Yan, S., Chen, H.L., 2002. A scheduling model and a solution algorithm for inter-city bus carriers. Transportation Research A 36, 805 825. Yan, S., Lin, C.G., 1997. Airline scheduling for the temporary closure of airports. Transportation Science 31, 7282. Yan, S., Shieh, C.W., Chen, M., 2002. A simulation framework for evaluating airport gate assignments. Transportation Research A 36, 885898.

Vous aimerez peut-être aussi