You are on page 1of 9

Transportation Research Part C 25 (2012) 113–121

Contents lists available at SciVerse ScienceDirect

Transportation Research Part C


journal homepage: www.elsevier.com/locate/trc

Robust optimization model of schedule design for a fixed bus route


Yadan Yan a,c, Qiang Meng b,⇑, Shuaian Wang b, Xiucheng Guo c
a
School of Civil Engineering, Zhengzhou University, Zhengzhou, Henan 450001, China
b
Department of Civil and Environmental Engineering, National University of Singapore, Singapore 117576, Singapore
c
School of Transportation, Southeast University, Nanjing, Jiangsu 210096, China

a r t i c l e i n f o a b s t r a c t

Article history: The time control point strategy is often adopted by bus operators in China and Singapore to
Received 27 November 2011 provide more reliable transit service. It is thus important to design a schedule, in which bus
Received in revised form 17 April 2012 drivers should devote their efforts to catch up a scheduled arrival time at a predetermined
Accepted 17 May 2012
time control point on a bus route because passengers can definitely benefit from a reliable
bus route schedule. This paper first proposes a novel reliable bus route schedule design
problem by taking into account the bus travel time uncertainty and the bus drivers’ sche-
Keywords:
dule recovery efforts. It proceeds to develop a robust optimization model for the proposed
Transit schedule
Time control point
problem, which aims to minimize the sum of the expected value of the random schedule
Monte Carlo deviation and its variability multiplied by a weighting value. A Monte Carlo simulation
Robust optimization model based solution method is subsequently designed to solve the robust optimization model.
Finally, a numerical example based on a real bus route in Suzhou city of China is carried
out to demonstrate the strength of the robust optimization model. We find that the optimal
scheduled travel time (or slack time) depends on bus drivers’ schedule recovery behavior
and on decision makers’ scheduling philosophies.
Ó 2012 Elsevier Ltd. All rights reserved.

1. Introduction

Public transport companies in China and Singapore often use the time control point strategy to design a schedule of a bus
route in order to provide more reliable transit service, in which bus drivers should try their best to catch up a scheduled
arrival time at a predetermined time control point (i.e., bus stop) on the bus route. There are a series of time control points
including the bus departure terminal on the bus route. Therefore, a schedule of the bus route consists of an arrival time at
each of these time control points (Wirasinghe, 2003). This schedule enables a bus driver to frequently adjust her/his bus
speed between two consecutive time control points. Since passengers can definitely benefit from a reliable transit schedule
(Peng et al., 2002), it is thus important to design a reliable bus route schedule.
Travel time between any two consecutive time control points on a bus route is largely determined by the corresponding
road traffic conditions. A bus driver, in reality, has limited flexibility to adjust her/his bus speed given the scheduled arrival
time at the next time control point. For example, a bus driver can hardly speed up on a stretch of very congested road seg-
ment. To reflect variability of the road traffic condition, it is rational to assume that travel time between two consecutive
time control points of a bus route is a continuous random variable with a given probability density function. The probability
density function of this random travel time can be actually calibrated or estimated from historical data gathered from the
automatic vehicle location (AVL) system. The AVL system is capable of gathering an enormous quantity and variety of oper-
ational, spatial, and temporal data (TCRP 113, 2006). This paper aims to determine a reliable schedule for a bus route by tak-
ing into account uncertainty of bus travel time and bus drivers’ schedule recovery efforts.

⇑ Corresponding author. Tel.: +65 6516 5494; fax: +65 6779 1635.
E-mail address: ceemq@nus.edu.sg (Q. Meng).

0968-090X/$ - see front matter Ó 2012 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.trc.2012.05.006
114 Y. Yan et al. / Transportation Research Part C 25 (2012) 113–121

1.1. Literature review

Bus route design is a strategic level decision problem with considerable efforts, for example, Baaj and Mahmassani (1990,
1991, 1992, 1995), Ceder (2007), Beltran et al. (2009) and Cipriani et al. (2012). Given a bus route, the schedule design of the
bus route is a operational level decision problem and it includes the following fundamental decisions: (1) determine bus dis-
patching frequencies or headways; (2) select some bus stops as time control points whose scheduled arrival times are to be
published as the schedule or serve as a reference (i.e., internal schedule) for both supervisors and drivers to determine if a
particular trip is running early or late; and (3) determine the scheduled arrival time for each time control point and recovery
time to schedule at the end of the bus route.
The first decision is relatively independent of the other two. It is mainly about how to meet the passenger demand along
the route with a balance consideration of both reasonable vehicle utilization and acceptable waiting times of boarding pas-
sengers. As for the second decision, a number of rules of thumb have been followed by transit operators (Wirasinghe and Liu,
1995). Major intersections that are widely recognized and possess good pedestrian amenities like sidewalks and actuated
traffic signals make good time control points. It is reasonable and practical to locate intermediate time control points at
major trip generation locations such as shopping centers, hospitals, and government buildings. Time control points are also
useful at locations where time is critical, such as major employment centers and intersecting bus routes or rail centers (TCRP
30, 1998).
In this study, we shall not be concerned with the first two decisions. Rather, we wish to examine the scheduled arrival
time for each time control point and recovery time to schedule at the end of the bus route. The time from a bus’s departure
at one terminal to its next departure in the reverse direction is called the half-cycle time. It is the sum of scheduled route
travel time and recovery time. The first step of the schedule design is to determine the half-cycle time based on the distri-
bution of route travel times. Because the purpose of recovery time is to limit the likelihood that delays encountered in one
trip will propagate to the next, the half-cycle time is based logically on a high-percentile value of route travel time. The 95th-
percentile route travel time is often set as the half-cycle time criterion, according to Tri-Met’s and Brussels’ transit agency’s
practices (TCRP 113, 2006). Then scheduling travel time on segments between adjacent time control points, or equivalently
setting arrival times at time control points relative to the trip start time follows. The scheduled arrival time for a time control
point is chosen by taking the mean travel time (including passenger processing time at intermediate bus stops between two
consecutive time control points and passenger processing time at the upstream time control point) between two time con-
trol points, combining with slack time and adding the sum to the scheduled arrival time of the upstream time control point.
The slack time is the difference between the scheduled and actual average travel times, i.e., the scheduled travel time on the
route segment between time control points is defined as the mean travel time plus slack time. Essentially, the schedule de-
sign problem we focus on is to determine the optimal slack time to be added to the mean travel time between two adjacent
time control points. Hence, the scheduled route travel time is the sum of scheduled travel times on all segments along the
bus transit route. Finally, the recovery time is set as the difference between half-cycle time and scheduled route travel time.
There are a number of studies on the transit schedule design. Carey (1994) carried out a comprehensive work on a sche-
dule-based transit system. He derived a group of integral equations to describe the arrival- and departure-time distributions
and formulated the objective function as a combination of travel cost and the cost of deviation from a schedule. Wirasinghe
and Liu (1995) developed an analytical model to determine the time control point location as well as amount of the slack
time in transit schedule design. Liu and Wirasinghe (2001) further proposed a simulation model to design a schedule for
a fixed transit route with the holding control strategy. Wirasinghe (2003) examined the expected total cost of travel time
and delay penalty affected by the choice of slack time for a special case of a single link (non-stop) route where passengers
board at a terminal and vehicles run non-stop to the destination. Dessouky et al. (1999) analyzed the real bus operation data
collected by the Los Angeles County–Metropolitan Transit Agency and suggested a slack ratio (slack time to scheduled travel
time) of 0.25 on the three bus routes converging in downtown Los Angeles, meaning that actual travel time over any segment
can be no less than 75% of the scheduled travel time. Zhao and Dessouky (2006) studied the problem of determining the opti-
mal slack time to be added to the mean round-trip travel time that minimizes the passengers’ expected waiting times. Furth
and Mull (2009) focused on schedules of long headway routes with time point holding and derived the strictness of a time
point and for dispatching reliability at the terminal, which were related to the amount of slack time within the travel time
schedule and within the scheduled layover. In contrast with the above schedule-based holding policies, Daganzo (2009),
Daganzo and Pilachowski (2011) and Xuan et al. (2011) have analyzed the headway-based dynamic holding policies to avoid
bus bunching.
Most of the above literature focused on the scheduled departure time coupled with the bus holding control strategy.
However, the bus holding control strategy is not allowed or unacceptable for some bus operators. For example, in Singapore,
a bus driver is reluctant to hold his bus because it means a higher possibility that they will finish the trip late; the bus oper-
ators resist holding because holding a bus at a bus stop is prohibited by the regulators since it causes congestion at the bus
stop and other buses are unable to use the bus stop for passenger processing. Holding also causes impatience and anxiety for
passengers. Generally, the objective functions of the models built in the literature were designed to minimize the mean cost
of waiting time for passengers or the mean total cost of waiting time, travel time and operating cost. This design may lead to
large variations in the objective values for different simulation runs or test runs, although the mean value is optimal. Besides,
a bus driver tends to adjust the speed between two consecutive time control points in order to ensure transit service reli-
ability. However, the existing studies neglected bus drivers’ schedule recovery efforts under the schedule-based operation,
Y. Yan et al. / Transportation Research Part C 25 (2012) 113–121 115

which was first investigated by Kalaputapu and Demetsky (1995) and has been considered as a control factor in the bus ar-
rival time prediction model (Lin and Bertini, 2004; Chen et al., 2005).

1.2. Objectives and contributions

The objective of this study is to develop a model to design a reliable schedule for a fixed bus route with a series of time
control points, taking into account the bus travel time uncertainty and bus drivers’ schedule recovery efforts. The above lit-
erature review clearly indicates that the reliable bus transit schedule design proposed in this study is a new research issue
with practical importance. We will first establish a robust optimization model for the reliable transit schedule design prob-
lem. The concept of remaining impact after recovery is proposed to describe the schedule deviation that cannot be recovered
after the bus driver’s adjustment. The objective function is constructed to minimize a sum of the expected value of the ran-
dom schedule deviation and its variability multiplied by a weighting value. Then a Monte Carlo simulation based solution
will be introduced to solve the model. To further assess applicability of the model and solution algorithm, a numerical exam-
ple from a bus route operated in Suzhou city of China will be used.

2. Assumptions and problem description

We first define the following notations before we describe the reliable bus schedule design problem:

Ti1,i travel time used by a bus from the time control point i  1 to the time control point i;
STi1,i scheduled travel time from the time control point i  1 to the time control point i;
T 0i1;i travel time from the time control point i  1 to the time control point i after the schedule recovery process;
SATi scheduled arrival time at the time control point i;
AATi actual arrival time at the time control point i;
si1,i slack time allocated on the segment from the time control point i  1 to the time control point i;
SDi schedule deviation at the time control point i.
Ai1,i driver’s adjustment on the segment from the time control point i  1 to the time control point i;
bi1;i adjustment factor on the segment from the time control point i  1 to the time control point i, 0 6 bi1;i 6 1;
RAi1 remaining impact after recovery of the schedule deviation at the time control point i  1;
c1 parameter to penalize the earlier arrival at a time control point;
c2 parameter to penalize the later arrival at a time control point;
k a given non-negative weighting value;
h the scheduled headway;
N number of buses deployed on the bus transit route;
C half-cycle time.

We now consider a typical bus route, shown in Fig. 1, with n stops which are numbered from 1 denoting one end to n
representing the other end. It is assumed that there are m stops ðm 6 nÞ with the arrival time requirement out of these n
stops, including the bus stop 1 (i.e., the bus departure terminal). Without loss of generality, the bus stop 1 is the first time
control point (i.e., time control point 1) and bus stop n is the last time control point (i.e., time control point m).
To reflect variability of the road traffic conditions, it is assumed that Ti1,i is a random variable with a known probability
density function which can be actually calibrated and estimated from historical AVL data. Estimation of the probability dis-
tribution of segment travel times should be under unscheduled condition, i.e., all travel times will be appropriate observa-
tions only when bus drivers have not intentionally ‘‘killed’’ time along the route. Notice that the definition of segment travel
time here includes the time for passenger processing (i.e., passenger boarding and alighting time). The AVL data collection
process leads to a sample of stop level observations. The stop level data include information related to the bus, the stop,
when the bus arrives at the stop, and when the bus leaves the stop. The segment travel time Ti1,i is calculated as the

Bus trip direction

`
Stop 1(Time Time control Time control Time control Stop n (Time
control point1) point i-1 Ti-1,i point i point j control point m)

Time control point

Bus stop

Fig. 1. A bus route with predetermined time control points.


116 Y. Yan et al. / Transportation Research Part C 25 (2012) 113–121

difference between the arrival time for the time control point i  1 and the arrival time for the time control point i. These
segment travel time observations should be under unscheduled condition, i.e., no slack time is present in the real data.
The scheduled travel time STi1,i from the time control point i  1 to the time control point i is defined as follows:
ST i1;i ¼ EðT i1;i Þ þ si1;i ð1Þ
where E(Ti1,i) is the expected value of Ti1,i.
Assuming that the arrival time SAT1 at bus stop 1 is given, the arrival time scheduled at the time control point i
(i = 2, 3, . . . , m) is determined by
SAT i ¼ SAT i1 þ ST i1;i ; i ¼ 2; 3; . . . ; m ð2Þ
Too little scheduled travel time designed for a particular segment between two consecutive time control points can frus-
trate bus drivers to catch the arrival schedule and create potential safety issues such as speeding, pushing traffic lights and
being abrupt with passengers. Too much scheduled travel time creates unnecessary travel time for passengers and the inef-
ficient utilization of bus vehicles and drivers. Inaccurate travel time also contributes to missed transfers, schedule reliability
complaints and time-consuming schedule corrections.
Besides, if a bus driver is informed by her/his supervisor the schedule deviation at the time control point i  1, she/he will
try to adjust her/his bus speed to keep the buses on schedule at time control point i as much as possible. Because of this
behavior response, schedule recovery process is frequently adopted during bus operation on the bus route. Hence, the
bus drivers’ schedule recovery behavior should also be taken into account during transit schedule design.

3. Robust optimization model

The schedule deviation associated with the deviation from the scheduled arrival time at a particular time control point
i = 2, 3, . . . , m can be calculated by
SDi ¼ SAT i  AAT i ð3Þ
If SDi > 0, the bus arrives earlier than the schedule. Otherwise, it arrives later than the scheduled arrival time at the time
control point i.
We assume that once a bus in operation deviates from its schedule at a time control point, the bus driver will adjust the
bus speed, by either speeding up or slowing down on the next segment, to ensure schedule adherence at the next time con-
trol point. As shown in Fig. 1, if a bus arrives earlier at the time control point i  1, the bus driver will tend to slow down on
the segment from the time control point i  1 to the time control point i to ensure schedule adherence at time control point i;
if the bus arrives later at the time control point i  1, the driver will tend to speed up on this segment. The driver’s adjust-
ment on the segment from the time control point i  1 to the time control point i, denoted by Ai1,i, is assumed to be pro-
portional to the schedule deviation at the time control point i  1, according to Lin and Bertini (2004). It can be thus
calculated by
Ai1;i ¼ bi1;i  ðSAT i1  AAT i1 Þ; i ¼ 2; 3; . . . ; m ð4Þ
where bi1;i ð0 6 bi1;i 6 1Þ is the adjustment factor on the segment from the time control point i  1 to the time control point
i. It is also assumed that this adjustment parameter is a random variable because it depends on distances and traffic condi-
tion. b1,2 is assumed to be zero, i.e., no recovery efforts is adopted on the first segment.
In order to facilitate model formulation while not missing major characteristics of the schedule recovery behavior, we
propose the concept of remaining impact after recovery to describe the schedule deviation that cannot be recovered after
the bus driver’s adjustment. The formulation of the remaining impact after recovery for the segment between time control
point i  1 and time control point i is:
RAi1 ¼ ð1  bi1;i Þ  ðSAT i1  AAT i1 Þ; i ¼ 2; 3; . . . ; m ð5Þ
In other words, buses can partially recover from their schedule deviation. According to Fig. 1, it is straightforward to esti-
mate the travel time from the time control point i  1 to the time control point i after the schedule recovery process, denoted
by T 0i1;i , as follows:
T 0i1;i ¼ T i1;i þ Ai1;i ; i ¼ 2; 3; . . . ; m ð6Þ

We thus have
AAT i ¼ AAT i1 þ T 0i1;i ð7Þ

According to Eqs. (4)–(7), Eq. (3) can be rewritten as follows:


SDi ¼ SAT i  AAT i ¼ SAT i1 þ ST i1;i  ðAAT i1 þ T i1;i þ Ai1;i Þ ¼ ðSAT i1  AAT i1 Þ þ ðST i1;i  T i1;i Þ  Ai1;i
¼ ðSAT i1  AAT i1 Þ þ ðST i1;i  T i1;i Þ  bi1;i  ðSAT i1  AAT i1 Þ
¼ ðST i1;i  T i1;i Þ þ ð1  bi1;i Þ  ðSAT i1  AAT i1 Þ ¼ ðST i1;i  T i1;i Þ þ ð1  bi1;i Þ  SDi1 ð8Þ
Y. Yan et al. / Transportation Research Part C 25 (2012) 113–121 117

According to Eq. (1), it follows that


SDi ¼ ½EðT i1;i Þ þ si1;i  T i1;i  þ ð1  bi1;i Þ  SDi1 ; i ¼ 2; 3; . . . ; m ð9Þ
Since buses always arrive at the stop 1 (i.e., time control point 1) on time for passenger boarding and then start a new run,
there is no schedule deviation at the time control point 1, namely,
SD1 ¼ 0 ð10Þ
Let T = (Ti1,i: i = 2, 3, . . . , m) and b = (bi1,i: i = 2, 3, . . . , m) denote two random vectors to group the random travel times
and the adjustment factors, respectively. Let s = (si1,i: i = 2, 3, . . . , m) be the vector of all the slack times. The schedule devi-
ation shown in Eq. (9) can be alternatively represented by SDi(s, T, b), i = 2, 3, . . . , m, to highlight its randomness.
For a schedule-based bus operation system, reliability is usually defined as the match between the scheduled plan and the
actual operation. On-time performance depends on how bus operation is able to match the transit schedule. If the slack time
used for schedule design is inappropriate, the actual bus arrival time will be different from the designed arrival time on the
schedule, and the reliability may be quite low. Therefore, the scheduled travel time (i.e., the mean travel time plus slack
time) should be chosen to minimize an objective function that reflects the above disutilities. Let c1 and c2 be two parameters
to penalize the earlier and later arrival at a time control point, respectively. The generalized schedule deviation at a time
control point can be formulated by
GSDi ðs; T; b; c1 ; c2 Þ ¼ c1 maxðSDi ðs; T; bÞ; 0Þ þ c2 maxðSDi ðs; T; bÞ; 0Þ ð11Þ
Values of the two parameters c1 and c2 involved in Eq. (11) are determined by bus operators according to their preferences.
Their impact on the solution will be evaluated in the section of numerical example. A reliable transit schedule can be thus
obtained by solving the following robust optimization model:
X
m
min FðsÞ ¼ ½EðGSDi ðs; T; b; c1 ; c2 ÞÞ þ k  EðjðjSDi ðs; T; bÞj  EðjSDi ðs; T; bÞjÞÞjÞ ð12Þ
i¼2

subject to
X
m
hN
si1;i 6 C ð13Þ
i¼2
2

where k is a given non-negative weighting value, h is the scheduled headway, N is the number of buses deployed on the bus
transit route and C is the half-cycle time.
Obviously, the proposed reliable transit schedule problem aims to determine a slack time solution so that the total sche-
dule deviation costs of all time control points are minimized. There are two terms in the bracket of the objective function
shown in Eq. (12). The first term is the expected value of the generalized schedule deviation at the time control point i.
The second term measures the absolute difference from the expected value of the random variable |SDi(s, T, b)|, called the
mean-absolute-deviation (Ben-Tal et al., 2009; Shapiro et al., 2009) and it also measures the variability of the random var-
iable. The weighting value k shown in Eq. (12) reflects importance of the model robustness (i.e., variability) and decision
makers’ risk aversion about schedule deviation variability. The constraint (13) implies that the total slack time that can
be allocated to each segment between adjacent time points for one-way bus operation along the route is not greater than
the value hN2
 C, according to Liu and Wirasinghe (2001). For example, if we take k ¼ 0, the induced model becomes a sto-
chastic programming problem without the model robustness or reliability of the bus route.
The penalty ratio (c1/c2) may be different for different scheduling philosophies. To evaluate impact of parameters c1 and
c2, we can further present the following two special cases of the robust optimization model by setting c1 = 2c2 and c1 = 0.5c2.
Bus operators usually have their preferred scheduling philosophy. Some operators suggest that earlier arrival is more harm-
ful to passengers than later arrival while others suggest that later arrival causes more inconvenience to passengers and extra
operating cost to them (Wirasinghe, 2003; TCRP 113, 2006). In the situation when c1 = c2, earlier arrival and later arrival are
assumed to be equivalently harmful by bus operators. When c1 = 2c2, earlier arrival is assumed to be more harmful than later
arrival by transit operators. The penalty of a unit of schedule deviation for earlier arrival is assumed twice as much as that of
a unit of schedule deviation for later arrival. When c1 = 0.5c2, later arrival is assumed to be more harmful than earlier arrival
by transit operators. The penalty of a unit of schedule deviation for later arrival is assumed twice as much as that of a unit of
schedule deviation for earlier arrival.
Besides, considering the special characteristics of schedule design problem, other constraints can be added as needed.
Examples of these constraints may read: the total slack time and the slack time allocated to each segment between adjacent
time control points can be handled as integers; no more than 2 min of slack time can be allocated to any segment except for
the last segment.

4. Monte Carlo simulation based solution method

The expected value term involved in the objective function of the robust optimization model shown in Eq. (12) does not
have a closed form. To solve the model, we first use the Monte Carlo simulation approach to generate K realizations of
118 Y. Yan et al. / Transportation Research Part C 25 (2012) 113–121

random variables (T, b), denoted by {(T(k), b(k)), k = 1, . . . , K}. The expected value can be approximated by their respective
sample mean based on GSDi(s, T(k), b(k), c1, c2) and |SDi(s, T(k), b(k))| with respect to a given realization (T(k), b(k)), namely

1X K  
EðGSDi ðs; T; b; c1 ; c2 ÞÞ  GSDi s; TðkÞ ; bðkÞ ; c1 ; c2 ð14Þ
K k¼1

K  
1X  ðkÞ 
EðjSDi ðs; T; bÞjÞ  SDi ðs; T ; bðkÞ Þ ð15Þ
K k¼1

 !
K    1X K   
1X   ðkÞ ðkÞ   ðkÞ ðkÞ  
EðjðjSDi ðs; T; bÞj  EðjSDi ðs; T; bÞjÞÞjÞ   SDi ðs; T ; b Þ  SDi ðs; T ; b Þ  ð16Þ
K k¼1  K k¼1 

With these sample means, the proposed robust optimization model can be approximated by the minimization model:
 !
m    1X K    
1X K X m
ðkÞ ðkÞ kX K X
  ðkÞ ðkÞ   ðkÞ ðkÞ  
min FðsÞ ¼ GSDi ðs; T ; b ; c1 ; c2 Þ þ  SDi ðs; T ; b Þ  SDi s; T ; b   ð17Þ
K k¼1 i¼2 K k¼1 i¼2  K k¼1 

subject to
X
m
hN
si1;i 6 C ð18Þ
i¼2
2

Since parameter k is a non-negative parameter, the minimization model expressed by Eqs. (17), (18) can be easily and
equivalently transformed into a mixed integer linear programming (LP) model by means of the integer programming model
building techniques (Nemhauser and Wolsey, 1988). It thus can be efficiently solved by any optimization solver such as
CPLEX. It should be pointed out the convergence of the Monte Carlo simulation based method has already been proved in
the literature when the sample size approaches infinity. Mak et al. (1999) gave statistical analysis results between the qual-
ity of the approximated solution and sample size. These results can guide us to take an appropriate sample size for a given
instance.

5. Numerical example

We take the 1W bus route shown in Fig. 2 in Suzhou city of China as a numerical example to assess the proposed robust
optimization model. This bus route is a heavily traveled cross-city route running from its northern terminal at Suzhou Rail
Station, through the old central district, to its southern terminal at New Village at Wuzhong district. It is 13.2 km long with

Fig. 2. 1W bus route and its four time control points.


Y. Yan et al. / Transportation Research Part C 25 (2012) 113–121 119

Table 1
Statistical analysis results (time unit: min).

Segment no. Stops included Segment length (km) Travel time


Mean Std. Median Minimum Maximum >Mean (%)
1 1–8 6.1 18.61 2.14 17.95 14.12 24.57 58
2 8–13 2.8 12.66 2.13 12.23 9.00 18.33 60
3 13–19 4.3 16.21 2.59 15.91 12.70 30.93 58

Std. denotes the standard deviation of segment travel times.


‘‘>Mean’’ denotes the percentage of segment travel times larger than the mean segment travel time.

Table 2
Lognormal distribution calibration for travel times on each segment.

Segment no. Estimated parameters for lognormal distribution Test value (A–D test)
Location Scale
1 2.9171 0.1117 0.5578
2 2.5254 0.1622 0.1968
3 2.7759 0.1363 0.8397

19 bus stops including the starting and ending terminal. The afternoon peak period (16:30–18:30) is considered in this
example. Currently, 1W bus route operates in an unscheduled and uncontrolled manner. The bus route at present does
not operate according to a schedule consisting of arrival times at some time control points. There is only a scheduled depar-
ture time from the southern terminal for the northbound direction trip and a scheduled arrival time at the southern terminal
for the southbound trip. Twenty five buses are currently deployed on the 1W bus route and all 25 buses are equipped with
GPS devices.
We assume that the northern terminal (i.e., stop 1), the stop 8, the stop 13 and southern terminal (i.e., the stop 19) are the
four time control points, shown in Fig. 2, which are numbered from time control point 1 to time control point 4. AVL data in
the afternoon peak period on May 4–6, 2011 are used for the following analysis. The mean total travel time along the route is
about 47.35 min, which gives an average commercial speed of 17 km/h. 95th-percentile of route travel times is 54.20 min
and it is taken as the half-cycle time, namely C = 54.2. Table 1 gives the statistical results for each of the three segments
based on the 60 AVL data sets.
The scheduled headway publicly published is 5 min. Hence, the total slack time that can be allocated is
hN 5  25
C ¼  54:2 ¼ 8:3 min
2 2
Based on the historical AVL data gathered from the 1W bus transit route, we find that the travel time on each segment
follows a lognormal distribution (shown in Table 2) because it statistically passes the Anderson–Darling (A–D) goodness-of-
fit. Regarding the adjustment factor, Chen et al. (2005) used the AVL data and the timetable from a reputable transit agency
in the northeastern United States and obtained the distribution of the adjustment factor along the route. Their results show
that on most segments, the average adjustment factors were mostly between 0.5 and 0.5. We assume the adjustment factor
is non-negative, i.e., bus drivers are always trying to recover the schedule deviation at the preceding time control point.
Whether the bus arrives at the time control point earlier or later lies on whether the schedule deviation at the time control
point is positive or negative. It is thus assumed that the two adjustment factors b2,3 and b3,4 are uniformly distributed in the
interval [0, 0.5]. We take the sample size K = 500. We set out three cases for three values of the penalty ratio (c1/c2), assuming
that k ¼ 0:3. When [Case-1] is applied, earlier arrival and later arrival are considered to be equivalently harmful. The three
cases are solved by CPLEX 12.1, and Table 3 gives the final results.
Results of [Case-1] (shown in Table 3) indicate negative optimal slack times for the segment travel times. This might be
counter-intuitive at first sight. However, it is reasonable because in contrast to most of the studies, we take into account
drivers’ recovery efforts. Moreover, we penalize both earlier arrival and later arrival, and we do not allow bus holding

Table 3
Results obtained from the three cases (time unit: min).

Segment 1 Segment 2 Segment 3


Slack time Scheduled Slack time (s2,3) Scheduled Slack time (s3,4) Scheduled
(s1,2) travel time (ST1,2) travel time (ST2,3) travel time (ST3,4)
[Case-1] ðc1 ¼ c2 ¼ 1; k ¼ 0:3Þ 0.15 18.46 0.11 12.55 0.07 16.14
[Case-2] ðc1 ¼ 2c2 ¼ 2; k ¼ 0:3Þ 0.93 17.68 0.56 12.10 0.29 15.92
[Case-3] ðc1 ¼ 0:5c2 ¼ 1; k ¼ 0:3Þ 0.60 19.21 0.34 13.00 0.21 16.42
120 Y. Yan et al. / Transportation Research Part C 25 (2012) 113–121

Table 4
Recovery time needed in the three cases (time unit: min).

[Case-1] [Case-2] [Case-3]


Total slack time 0.33 1.78 1.15
Scheduled route travel time 47.15 45.70 48.63
Recovery time 7.05 8.50 5.57

Table 5
Example of converting calculating results into bus operation schedule.

Scheduled arrival Scheduled arrival Scheduled arrival time


time at stop 8 time at stop 13 at southern terminal
(stop 19)
[Case-1] 16:57:28 17:10:01 17:26:09
[Case-2] 16:56:41 17:08:47 17:24:42
[Case-3] 16:58:13 17:11:13 17:27:38

Fig. 3. Results of sensitivity analysis of the weighting value k.

due to reasons mentioned in the introduction section. As a consequence, it is possible, and it is indeed the case for the 1W
bus route, that negative slack times should be scheduled.
[Case-2] assumes that earlier arrival is more harmful. Hence, shorter segment travel times should be scheduled so that the
chance of earlier arrival is lower than the chance of later arrival. Contrary to [Case-2], more scheduled segment travel times
are needed in [Case-3]. As a result, the chance of later arrival is lower than the chance of earlier arrival in [Case-3].
Table 4 also shows the recovery time planned in three cases. The calculating results can be converted into the schedule for
bus operation. For example, assuming that a bus’s scheduled arrival time at the northern terminal is 16:39:00, its scheduled
arrival times at other three time control points are shown in Table 5.
The effect of the weighting value k on the optimal slack times is analyzed by varying the weighting values k from 0 to 20,
as shown in Fig. 3. Fig. 3 illustrates that for the given values of c1 and c2, the slack times change significantly when k P 5.
When k 6 1, in other words, when k is smaller or equal to the values of penalty for the earlier or later arrival c1 and c2, the
slack times do not change significantly with k.
From the above discussions, it can be concluded that if the decision maker wishes to have robust schedule by using the
proposed model, he/she should choose values of c1 and c2 which reflect his/her preferred scheduling philosophies and a va-
lue of k which reflects his/her risk aversion about schedule deviation variability.
Y. Yan et al. / Transportation Research Part C 25 (2012) 113–121 121

6. Conclusions

In bus operations, schedule is of significance for bus operators to provide reliable service. This paper has investigated the
schedule design problem for a fixed bus route without holding control. Bus drivers’ schedule recovery behavior is taken into
account. Different operators’ scheduling preferred philosophies are discussed. With these practical considerations, the model
proposed can be employed as a useful and practical quantitative analysis tool to assist the schedule design for transit oper-
ators. The scheduled arrival time at a time control point is determined by its arrival time at the upstream time control point
and the scheduled travel time on the segment between these two time control points, while the scheduled segment travel
time is defined as the sum of mean segment travel time and slack time. The proposed model has been evaluated by a case
study based on a practical bus route. Future research for this study may focus on investigating the penalty for a unit of sche-
dule deviation and conducting a comparison study with other schedule design methodologies.

Acknowledgements

The authors wish to thank three anonymous referees for their helpful comments and valuable suggestions which consid-
erably improved the exposition of this work.

References

Baaj, M.H., Mahmassani, H.S., 1990. TRUST: a LISP program for the analysis of transit route configurations. Transportation Research Record (1283), 125–135.
Baaj, M.H., Mahmassani, H.S., 1991. An AI-based approach for transit route system planning and design. Journal of Advanced Transportation 25 (2), 187–
209.
Baaj, M.H., Mahmassani, H.S., 1992. Artificial intelligence-based system representation and search procedures for transit route network design.
Transportation Research Record (1358), 67–70.
Baaj, M.H., Mahmassani, H.S., 1995. A hybrid route generation heuristic algorithm for the design of transit networks. Transportation Research Part C 3 (1),
31–50.
Beltran, B., Carrese, S., Cipriani, E., Petrelli, M., 2009. Transit network design with allocation of green vehicles: a genetic algorithm approach. Transportation
Research Part C 17 (5), 475–483.
Ben-Tal, A., Ghaoui, L.E., Nemirovski, A., 2009. Robust Optimization. Princeton University Press, Princeton.
Carey, M., 1994. Reliability of interconnected scheduled services. European Journal of Operational Research 79, 51–72.
Ceder, A., 2007. Public Transit Planning and Operation. Elsevier, Amsterdam.
Chen, M., Liu, X.B., Xia, J.X., 2005. Dynamic prediction method with schedule recovery impact for bus arrival time. Transportation Research Record (1923),
208–217.
Cipriani, E., Gori, S., Petrelli, M., 2012. Transit network design: a procedure and an application to a large urban area. Transportation Research Part C 20 (1), 3–
14.
Daganzo, C.F., Pilachowski, J., 2011. Reducing bunching with bus-to-bus cooperation. Transportation Research Part B 45, 267–277.
Daganzo, C.F., 2009. A headway-based approach to eliminate bus bunching: systematic analysis and comparisons. Transportation Research Part B 43, 913–
921.
Dessouky, M., Hall, R., Nowroozi, A., Mourikas, K., 1999. Bus dispatching at timed transfer transit stations using bus tracking technology. Transportation
Research Part C 7, 187–208.
Furth, P.G., Mull, T.H.J., 2009. Optimality conditions for public transport schedules with timepoint holding. Public Transport 1, 87–102.
Kalaputapu, R., Demetsky, M.J., 1995. Modeling schedule deviations of buses using automatic vehicle-location data and artificial neural networks.
Transportation Research Record (1497), 44–52.
Lin, W.H., Bertini, R.L., 2004. Modeling schedule recovery processes in transportations for bus arrival time prediction. Journal of Advanced Transportation 38
(3), 347–365.
Liu, G., Wirasinghe, S.C., 2001. A simulation model of reliable schedule design for a fixed transit route. Journal of Advanced Transportation 35 (2), 145–174.
Mak, W.K., Morton, D.P., Wood, R.K., 1999. Monte Carlo bounding techniques for determining solution quality in stochastic programs. Operations Research
Letters 24 (1), 47–56.
Nemhauser, G.L., Wolsey, L.A., 1988. Integer and Combinatorial Optimization. John Wiley & Sons, Inc.
Peng, Z.R., Yu, D.L., Beimborn, E., 2002. Transit user perception of the benefits of automatic vehicle location. Transportation Research Record 4067, 127–133.
Shapiro, A., Dentcheva, D., Ruszczyński, A., 2009. Lectures on Stochastic Programming: Modeling and Theory. The Society for Industrial and Applied
Mathematics and the Mathematical Programming Society.
Transit Scheduling: Basic and Advanced Manuals, 1998. TCRP Synthesis of Transit Practice, 30, Transportation Research Board of the National Research
Council, Washington, DC.
Using archived AVL-APC Data to Improve Transit Performance and Management, 2006. TCRP Synthesis of Transit Practice 113, Transportation Research
Board of the National Research Council, Washington, DC.
Wirasinghe, S.C., Liu, G., 1995. Determination of the number and locations of time-points in transit schedule design. Annuals of Operations Research 60,
161–191.
Wirasinghe, S.C., 2003. Initial planning for urban transit systems. In: Lam, William H.K., Bell, Michael G.H. (Eds.), Advanced Modeling for Transit Operations
and Service Planning. Elsevier Science Ltd., UK.
Xuan, Y., Argote, J., Daganzo, C.F., 2011. Dynamic bus holding strategies for schedule reliability: optimal linear control and performance analysis.
Transportation Research Part B. http://dx.doi.org/10.1016/j.trb.2011.07.009.
Zhao, J.M., Dessouky, M., 2006. Optimal slack time for schedule-based transit operations. Transportation Science 40 (4), 59–539.

You might also like