Transportation Research Part B: Saeed Asadi Bagloee, Avishai (Avi) Ceder

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 18

Transportation Research Part B 45 (2011) 1787–1804

Contents lists available at ScienceDirect

Transportation Research Part B


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

Transit-network design methodology for actual-size road networks


Saeed Asadi Bagloee a,1, Avishai (Avi) Ceder b,⇑
a
PARSONS, P.O. Box 9123, Dubai, United Arab Emirates
b
Department of Civil and Environmental Engineering, University of Auckland, New Zealand

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

Article history: The main purpose of this study is to design a transit network of routes for handling actual-
Received 17 December 2010 size road networks. This transit-network design problem is known to be complex and
Received in revised form 12 July 2011 cumbersome. Thus, a heuristic methodology is proposed, taking into account the major
Accepted 12 July 2011
concerns of transit authorities such as budget constraints, level-of-service standards and
the attractiveness of the transit routes. In addition, this approach considers other impor-
tant aspects of the problem including categorization of stops, multiclass of transit vehicles,
Keywords:
hierarchy planning, system capacity (which has been largely ignored in past studies) and
Public transportation
Transit-network design
the integration between route-design and frequency-setting analyses. The process devel-
Newton gravity oped starts with the construction of a set of potential stops using a clustering concept.
Routes hierarchy Then, by the use of Newton gravity theory and a special shortest-path procedure, a set
Rail system of candidate routes is formed, categorized by hierarchy (mass, feeder, local routes). In
the last step of the process a metaheuristic search engine is launched over the candidate
routes, incorporating budgetary constraints, until a good solution is found. The algorithm
was tested on the actual-size transit network of the city of Winnipeg; the results show that
under the same conditions (budget and constraints) the proposed set of routes resulted in a
reduction of 14% of total travel time compared to the existing transit network. In addition
the methodology developed is compared favorably with other studies using the transit
network of Mandl benchmark. The generality of the methodology was tested on the recent
real dataset (pertaining to the year 2010) of the larger city of Chicago, in which a more
efficient and optimized scheme was proposed for the existing rail system.
 2011 Elsevier Ltd. All rights reserved.

1. Introduction

Many public transit agencies are using route networks which have not been reappraised for anywhere from 20 to 50
years. This provides sufficient motivation to seek an efficient network route-design method, based on certain objective func-
tions and a set of constraints. There are two main approaches to restructuring public transit routes: (1) at the route level or
for a small group of routes; (2) at the network level. For the first approach, Ceder (2007) suggested that an appropriate
restructuring approach was to simplify routes, accommodate new travel patterns, ease or eliminate transfers, reduce route
circuitry, or otherwise alter route configuration. The present work proffers a practical solution utilizing the second approach.
It is known (Ceder, 2007; Ceder et al., 2002; Ceder and Wilson, 1986) that a cost-effective transit operations planning process
is comprised of five efficient components: (i) designing the network of routes, (ii) frequency setting, (iii) timetabling, (iv)
vehicle scheduling and (v) crew scheduling and rostering. The first two, usually termed strategic planning (Guihaire and
Hao, 2008), are addressed in this study.

⇑ Corresponding author. Tel.: +64 9 3652807; fax: +64 9 3652808.


E-mail addresses: Saeed.Asadi@parsons.com (S.A. Bagloee), a.ceder@auckland.ac.nz (A. Ceder).
1
Tel.: +971 50 945 3192; fax: +971 4 336 7920.

0191-2615/$ - see front matter  2011 Elsevier Ltd. All rights reserved.
doi:10.1016/j.trb.2011.07.005
1788 S.A. Bagloee, A. Ceder / Transportation Research Part B 45 (2011) 1787–1804

Nomenclature

G(V, A) a network, with V and A, sets of nodes and links, respectively


Z set of traffic zones represented by a centroid
dp,q transit demand from zone p to zone q, and amounts of originating demand from and incoming demand to des-
P P p;q P P p;q
tination zone z 2 Z are: Oz ¼ p¼z 8q2Z d ; Dz ¼ 8p2Z q¼z d , respectively
‘(p, q) shortest (walking) distance between two nodes or stops or zones (p, q) in km
t(p, q) travel time on the shortest path between twonodes or stops or zones (p, q) in minutes
ti,j travel time in minutes on link (i, j) 2 A
q is a uniform random number between 0 and 1, with mean value of 0.5, that is q – q
Wv weight of node v 2 V
S0 set of Sieved stops (S0  V)
S set of candidate stops (S  S0 )
SL0 descended sorted S by Wv, v 2 S
b balancing factor for converting demand from zonal-level to candidate-stop level
m;n m;n
dc ðbÞ or dc demand on candidate-stop level (denoted by c), for (m, n) 2 S  S, function of b
R set of generated candidate routes
r a candidate-route with terminals (candidate-stops) s, e 2 S also denoted by r ¼ hs; ei 2 R
_ _
Ms ; Me mass (demand-related) at terminals (candidate-stops) s, e 2 S
f(s, e) friction (distance-related) between terminals (candidate-stops) s, e 2 S
~s;e
d fractional-maximal demand between terminals (candidate-stops) s, e 2 S
c

k(m) number of contribution of m 2 S in R as terminal ðhm; ni 2 RÞ


r(s, e) biased-length-index of terminals (candidate-stops) s, e 2 S
d stochastically desired length of passengers’ trips
y cumulative transit trips over total demand, 0 6 y 6 1
~t p;q is the manipulated travel time on link (p, q) 2 A – see Note below of travel time
sp,q free-transit-flow of link (p, q) 2 A
hp,q attraction index of link (p, q) 2 A
a, b parameters to regulate the impacts of hp,q, sp,q
DSPi: amount of directly served passenger for candidate-route generated at iteration i
DSPi decremented threshold value for qualification of candidate-route generated at iteration i
i1, i2, i3, (1 < i1 < i2 < i3) maximum-iteration number delineates hierarchy types of mass, feeder and local, respectively,
during candidate-route generations
~tðs; eÞ geometric mean of terminals’ travel time of 2-way route r ¼ hs; ei 2 R – see Note below of travel time
p(i) gradual-incremental term (0.5 6 p(i) 6 1) at iteration iused in Eq. (20)
ais P number ofP accesses/egresses provided to terminal (candidate-stop) s 2 Sup to iteration i
v ole ¼ ðp;eÞ2A sp;q þ ðe;qÞ2A sp;q free transit passing-through volumes at terminal (candidate-stop) e 2 S
trhs, ei travel time of 2-way path of route hs, ei – see Note below of travel time
X set of transit classes e.g., minibus, bus, metro
g denotes a transit class g 2 X
1(g) total capacity of transit vehicle class g in terms of number of passengers
pir collected point of candidate-route r 2 R of iteration i
B limited budget interpreted as the maximum available capacity (number of seats)
i
br load (or total boarding passengers) of r 2 R in iteration i
N total number of iterations in Route Selection and Frequency Setting (RSFS)
N0 total number of iterations in Phase 0 of RSFS where a space is generated using a set of stochastic scenarios to
shape the solution
Si scenario corresponding to iteration 1 6 i 6 N
cir updated required capacity for a candidate-route r 2 Rat the end of iteration 1 6 i 6 N
i
cSr needed capacity for candidate-rout r 2 R in the scenario of iteration i,r 2 Si
l^f i ; lf i
r r maximum and average load-factor for candidate-route r 2 R in the scenario of iteration i, r 2 Si, respectively
i;ðgÞ
fr number of needed fleet of transit vehicles class g 2 X for candidate-route r 2 R in the scenario of iteration i, r 2 Si
ðgÞ
hdwyr latest updated needed headway of candidate-route r with respect to transit class g 2 X
i i
v ehr latest updated allocation of vehicle class type ðv ehr ¼ gÞ to candidate-route r
CPir collective points of candidate-route r computed for iteration N0 < i 6 N.
S.A. Bagloee, A. Ceder / Transportation Research Part B 45 (2011) 1787–1804 1789

Note: Because the focus of this study is on transit-route design, the use of ‘‘travel time’’ refers to ‘‘transit travel time’’.
Wherever the network under consideration consists of automobile roadways, then certainly ‘‘travel time’’ is sub-
ject to both auto travel time and the transit system. For a separate transit system such as a rail network the tra-
vel time depends only on this transit network. Furthermore the ‘‘overall travel time’’ is comprised of cruising
time and dwelling time at stops considered in the transit assignment modules. Therefore, wherever travel time
is associated with transit assignment modules it implies overall transit travel time; otherwise it depends only on
a separated-system cruising time.

Objective functions of transit network design and frequency setting naturally involve the following considerations:

 Operator cost (Ceder, 2007).


 Fleet size (Ceder, 2007).
 Passenger waiting-time and travel-time (Bielli et al., 2002; Tom and Mohan, 2003; Fan and Machemehl, 2006b; Pattnaik
et al., 1998; Carrese and Gori, 2002; Shih et al., 1998; Salzborn, 1980; Silman et al., 1974; Xiong and Schneider, 1993).
 Number of transfers known (Ceder, 2007; Ceder and Wilson, 1986; Rapp and Gehner, 1976).
 Number of direct trips (Ceder, 2007).
 Transfer time (Baaj and Mahmassani, 1991; Baaj and Mahmassani, 1995).
 Profit (Yan and Chen, 2002).

User and operator perspectives are commonly used in the search for a transit network. Users seek a better level-of-service
and the operator needs to provide the service for the minimum possible cost. Studies report attempts to look for an optimal
balanced solution between improved level-of-service and minimum service cost. However, in practice, the local authority (or
government) is constrained by a limited budget which does not always allow the implementation of the identified optimal
transit network and service design. Thus, this limited budget should enter the equations in an effort to be both practical and
optimal. This direction is adopted by this study.
In the design of a transit network, the following process is usually followed in the literature. First, a route-generating
strategy is established to determine a set of candidate-routes. Second, a procedure is developed to combine the candi-
date-routes to generate an initial solution. Third, by use of an iterative process, based on the previously generated solutions,
a new solution is generated. This new solution is then tested and analyzed to identify the right direction for the next solution
until a satisfactory set of routes is created. Ceder and Wilson (1986) deconstructed the process into a sequence of compo-
nents. Ideally, all the components should be treated simultaneously, complete with interaction and feedback, but this is ex-
tremely complex and difficult; even the treatment of the components separately is already cumbersome or NP-hard
(Magnanti and Wong, 1984).
From the perspective of solution strategies in solving transit network design problems, there are usually two approaches:
exact search methods (exact optimization) and non-exact methods (heuristic). The computational complexity of exact opti-
mization methods are unable to address actual-size of road networks as shown in a number of articles (Salzborn, 1980; Yan
and Chen, 2002; Guan et al., 2003; Hasselström, 1981; Van Nes et al., 1988; Furth and Rahbee, 2000; Murray, 2003). At the
same time, the heuristic methods, freed from the constraint of finding exact solutions, are able to tackle actual-size prob-
lems. These types of solutions are covered extensively in the literature (Ceder, 2007; Ceder and Wilson, 1986; Bielli et al.,
2002; Tom and Mohan, 2003; Fan and Machemehl, 2006b; Pattnaik et al., 1998; Carrese and Gori, 2002; Shih et al., 1998;
Silman et al., 1974; Xiong and Schneider, 1993; Rapp and Gehner, 1976; Baaj and Mahmassani, 1991; Baaj and Mahmassani,
1995; Zhao, 2006; Ceder et al., 2002; Carrese and Gori, 2002; Gao et al., 2003). Consequently, a meta-heuristic solution algo-
rithm based on the Ant System (AS) as hybridized by a Genetic Algorithm (GA) is developed in this study.
Various aspects of the transit-network problem have been dealt with in the literature as follows.

 Practicality – being able to tackle actual-size problems (Bielli et al., 2002; Carrese and Gori, 2002; Silman et al., 1974;
Rapp and Gehner, 1976; Yan and Chen, 2002; Guan et al., 2003; Zhao, 2006; Ceder et al., 2002).
 Integrated planning – synchronizing routes and frequency setting (Tom and Mohan, 2003; Pattnaik et al., 1998;
Hasselström, 1981; Van Nes et al., 1988).
 Multiclass – design of a system containing different transit modes (e.g., metro, bus, ferry) and different route hierarchies
(e.g., collector, feeder, mass (or dominant)-route) in a systematic manner; apart from Carrese and Gori (2002) who con-
sidered three-route hierarchies the other studies focus on specific linkages such as feeder routes for a rail network and
alike (Salzborn, 1980; Knoppers and Muller, 1995).
 Transit assignment – a transit assignment method is required for a given solution; it is noted that in transit-network
design most researchers, because of computational complexities, use simplifications.
 Stop positioning – despite the importance of positioning the stops on the routes, only a handful of research papers con-
sider it and these are mostly concerned with the issue of optimal stop-spacing by the use of rules-of-thumb (Furth and
Rahbee, 2000; Saka, 2001). A notable study by Murray (2003) considers two variations: relocation of the stops on an exist-
ing transit network, and the optimal location of stops in a process to create or extend the network. However, interaction
between transit routes and stops is ignored.
1790 S.A. Bagloee, A. Ceder / Transportation Research Part B 45 (2011) 1787–1804

The primary focus of the current study is to provide systematic transit-planning tools for actual-size networks with the
consideration of the above aspects that were not treated collectively in past studies. The key aspects of this study are;

 We set our main target to tackle the actual-size cases.


 In the proposed methodology the algorithm is launched with an intuitive algorithm based on the concept of clustering to
identify the best set of candidate-stops.
 We synchronize the process of laying down the routes with setting the frequencies.
 Our proposed methodology is able to take different transit modes (rail, bus . . .) and hierarchies (local, feeder, mass) into
consideration.
 We employ two essential types of transit assignment methods in the process of planning: capacity-free and capacity-con-
straint transit assignment.

The structure of this study is as follows. Section 2 describes the principal concepts of the proposed methodology to design
optimal transit network. Section 3 provides details of the methodology. Section 4 investigates the case studies of Winnipeg
city, the Mandle benchmark and the recent real dataset of the city of Chicago and evaluates the methodology proposed. Sec-
tion 5 is the conclusion.

2. Concepts of the proposed algorithm

Given a road network, an exogenous fixed transit demand and a limited budget, the problem is to derive a comprehensive
transit plan (including routes and assigned vehicles) so that the total passengers’ discomfort is minimized. A route is com-
prised of two end points (terminals), a sequence of links, a timetable based on frequency setting and is 2-way (bidirectional).
The algorithm allows different and even distant paths (depending upon demand concentration) for either direction (see Eqs.
(7)–(13)), thus loop routes can also be considered. In any transit plan (scenario), the discomfort index of passengers from
origin p to destination qis designated as generalized time GTp,q:
GT p;q ¼ InVehTimp;q þ x1  WalkTimp;q þ x2  WaitTimp;q þ x3  Transfer p;q ð1Þ

where InVehTimp,q, WalkTimp,q, WaitTimp,q, Transferp,q are in-vehicle time, access/egress-walking time, waiting time, and
number of transfers or boarding, respectively, to get from p to q.
x1, x2, x3 > 0 are weights for waiting time, walking time and transfer/boarding penalty while in-vehicle’s weight is
assumed to be 1.
The objective function (U) is established as the total saved generalized time with respect to no-transit-plan scenario:
X X
max Ui ¼ GT 0p;q  GT ip;q ð2Þ
p;q p;q

where 0 and i denote no-transit-plan and scenario number, respectively. This formulation makes the definition of the objec-
tive function clearer in terms of discomfort index interpreted as generalized times, and emphasizes the catering to captive
users who have no other travel choice than transit.
The proposed algorithm consists of three components. In component 1 the location of stops is identified based on close-
ness to highly concentrated transit demand places and clustering factor. Component 2 is inspired by the Newton gravity the-
ory, to generate a set of candidate routes. In component 3 a search algorithm is run through the generated candidate-routes
to seek a good solution; this search procedure is encoded by means of a genetic algorithm (GA) and equipped with ant-sys-
tem (AS) collective points as the search engine. The GA is a search heuristic that imitates the process of natural evolution. In a
genetic algorithm, a population of strings (called chromosomes), encode candidate solutions to an optimization problem and
evolve toward better solutions. The AS is a heuristic search technique which seeks an optimal path in a graph, inspired by the
behavior of ants seeking a path between their ‘‘colony’’ and a ‘‘source of food’’. In this study the concept of collective points
associated with each ‘‘source of food’’ is exploited.
Fig. 1 outlines the developed methodology with the indication of the components. It consists of two parts (i) route gen-
eration and (ii) route selection and frequency setting.

3. Proposed algorithm

The algorithm proposed is based on the parameters and variables associated with network analysis, GA and AS proce-
dures and thus require the use of many notations. Following is the complete list of notations used throughout this study.

3.1. Categorization of stops

For each node a weight index is introduced for computing the closeness of the node to transit demand concentrated
places, known as transit centers. Then the nodes are sieved and those with a poor index are discarded. The remaining nodes
are then clustered. Finally the weightiest node of each cluster is labeled as a candidate-stop.
S.A. Bagloee, A. Ceder / Transportation Research Part B 45 (2011) 1787–1804 1791

Stop Category
Clustering concept

Calculate demand at
stop-level
Part 1: Route Generation

Route Route Generation module


detouring filter -Selecting terminals; Gravity model
Not more than -Laying down the routes on manipulated shortest-path
1.57 -Hierarchies: mass-feeder-local

No
∀s ∈ S has a min ?

Yes

Part 2: Routes Selection and Frequency Setting

Scenario
Phase 0: Stochastic Generation Evaluation
Random selection subject to budget Capacity-free
assignment

Updating No
Points ∀r ∈ ℜ counted ?
needed capacity,
headway, Yes
vehicle type

Scenario
Evaluation
Route Phase 1: Sophisticated Generation Capacity-
selection filter Selection upon collective point constrained
Headway filter assignment

No
Impossible to
improve?

Yes

A Good Solution

Fig. 1. Proposed methodology for the transit-network design problem.

3.1.1. Computing weight of nodes


The measure of the closeness of nodes to transit centers is called as the nodes’ weight, and takes on the following
formulation:

X Oz þ Dz
Wv ¼ ‘ðv ;zÞ
for v 2 V; z for which ‘ðn; zÞ 6 1:0 ð3Þ
z e 0:3

This definition has significant implications: Passengers (Oz + Dz) choose the stops which are within their walking distance.
They do not tolerate a long walking distance (e.g., if a walking distance is multiplied, their dissatisfaction will be more than
twice as much, thus the exponential perception of distance is used (e())). Passengers’ satisfaction lies within a certain stop
space (assumed minimum 0.3 km in the denominator of the power). From 0.3 km up to 1 km (maximum walking distance,
‘(n,z) 6 1.0), the contribution of the transit centers (Oz + Dz) to the weight of the node rapidly diminishes no matter how large
the centers are.
1792 S.A. Bagloee, A. Ceder / Transportation Research Part B 45 (2011) 1787–1804

3.1.2. Sieving nodes based on their weights


Set of sieved node (S0  V) is derived as:

S0 ¼ f v 2 V v for which W v P 1=eg ð4Þ


The threshold (1/e) means that any node in the vicinity of a satisfactory walking distance (independent of the number of
passengers) is considered as a possible stop which is a sieved-node (see Eq. (3) assuming that Oz + Dz = 1, ‘(n, z) = 0.3).

3.1.3. Clustering and extracting candidate stops


The sieved set of nodes might have bunches which are too close to a sieved-node; this may breach the assumption of min-
imum stop distance. On the basis of common sense we assumed 300 m as minimum stop distance (0.3 km). Thus the sieved-
nodes in the same vicinity are bundled as a cluster such that the distance between adjacent clusters would not be less than
0.3 km. Then from S  S0 which is a set of candidate stops, the weightiest sieved-node in each cluster is obtained as follows.

– Step 0 – Preparation
 Initialize set of candidate-stops S
 Sort set S0 by sieved-nodes’ weight (Wv,v 2 S) in descending order to get sorted-list SL0
– Step 1 – Clustering
– Remove the weightiest node from SL0 as s⁄; If s⁄ does not fall within 0.3 km distance to any of the already identified
candidate-stop then: s⁄ ? S
– Step 2 – Ending
 If SL0 – ; then repeat Step 1
 End.

The outcome of this module is a set of candidate stops, using a coverage perspective, on which candidate routes are gen-
erated, and from which a good scenario of candidate-routes emerges. Thus, some poorly located candidate stops would not
be considered in the final scenario because of the issue of route efficiency. This, and the topology of the candidate stops, im-
plies that the stops would not be as close as 0.3 km on all the routes.
During the route-generation process, without engaging initially the transit-assignment part (to minimize computational
time), a preliminarily route-evaluation procedure is designed. Given a transit demand figure between stops, simple calcula-
tions can produce the overall passenger load of the route (see Section 3.2.3). However, at the end of each iteration, a proper
transit-assignment procedure is executed). The next section computes the transit demand at the stop-level (demand from
stop to stop).

3.1.4. Demand at the stop-level


A fast and initial calculation for the preliminary evaluation of scenarios (sets of routes) is introduced, in which direct de-
mand interaction between stops is sought. This adopts a heuristic approach in which the stops’ closeness to the zones and
the (corresponding) transit demand exchanging amounts are utilized as follows (this implicitly means converting demand
from zonal level to stop level):
X p;q
m;n d
dc ðbÞ ¼ b  ‘ðm;pÞ ‘ðn;qÞ
for p; q 2 Z; m; n 2 S; ‘ðm; pÞ 6 1:0; ‘ðn; qÞ 6 1:0 ð5Þ
e 0:3  e 0:3
m;n
where b is a balancing factor to ensure that summation of the converted demand ðdc ðbÞÞ is equal to the original dp,q. This
factor is simply identified as:
,
XX p;q
XX m;n
b¼ d dc ð1Þ ð6Þ
p2Z q2Z m2Z n2
m;n m;n
Following the b computation, and for simplicity, we refer to dc ðbÞ as dc .

3.2. Route generation

In an iterative-convergence process, two terminals (s, e) 2 S  S and then the itinerary of candidate routes are identified to
generate a set of candidate-route ðRÞ. Candidate routes are denoted by r ¼ hs; ei 2 R.

3.2.1. Selecting terminals


It is likely to lay down a route between two terminals (end points of a route) which will result in high passenger demand
subject to elements like distances and available required facilities. This likelihood opens the window for the use of Newton–
Gravity theory in which total demand of each terminal and distance can be used as ‘‘mass’’ and ‘‘distance’’ of gravity. A grav-
ity-index (GIs,e) between terminals s, e 2 S is defined. Possible locations of terminals such as available facilities in place
(mainly designated in the outskirts of the cities) can easily be implemented by revising the setS, under which routes with
low demands at terminals, but heavy passenger loads otherwise, may be captured too. The higher the gravity-index the more
chance for r = hs, ei to get into R which can be described as follows.
S.A. Bagloee, A. Ceder / Transportation Research Part B 45 (2011) 1787–1804 1793

 _ _ q
Ms  Me
GIs;e ¼ ~s;e  qkðsÞ  qkðeÞ  qrðs;eÞ
d where s; e 2 S; hs; ei R R; ‘ðs; eÞ > 1:8; where ð7Þ
f ðs; eÞ

_ _
 M s ; M e and f(s, e) refer to some sort of demand and distance called mass and friction respectively; more masses and less
friction means more gravity. These arguments delineate candidate-route hierarchy type (mass, feeder, local). Definition of
transit route hierarchy is mainly based upon the routes’ ridership as well as their length (Ceder, 2007). For instance higher
ridership on longer routes intuitively is known as a mass route whereas routes with low demand and shorter length cater-
ing to the peripheral areas are known as local routes. In between the feeder routes connecting local routes to the mass
routes exist. We shall elaborate and quantify the above definitions later in Section 3.2.3.
 q (a random variable) gives a stochastic chance to every pair of candidate-stop to be chosen as terminals.
~s;e ¼ 1 þ ds;e = maxðdm;n Þ
d ð8Þ
c c c
m;n2S

16d ~s;e 6 2 called fractional-maximal demand which provides a higher chance to those terminals with a high direct de-
c
mand (maximum 2); for terminals with no direct demand (e.g., heavy volumes in between) the chance still exists (min-
imum 1).
 k(s) indicates how many times candidate stop "s 2 S has been selected as a terminal; thus qk(s), qk(e) give less chances to
those who have been selected already; this aims to exploit a wide range of solution space.
 qr(s, e); a penalty term; aims to select terminals of around a stochastically desired length (d) is embedded in r(s, e) as follows:
 
tðs; eÞ  d
rðs; eÞ ¼  
 ð9Þ
d
r(s, e) is called biased-length-index of s, e, until terminals’ travel time reaches the desired length (t(s, e) ? d),
thusr(s, e) ? 0 and qr(s, e) ? 1 (no penalty). d is derived from the overall tendency, or desire, of passengers in terms of
trip length to support the following notion: if the majority of the demand is associated with long trips, designing short-
or medium-length routes does not make sense.
Extracting the desired-length (d) of trip is based on depicting accumulated transit trips versus travel distance reminiscent
of S-shape trends (or in general a logistic model). In addition, the most common models for trip distribution are gravity-
based models usually associated with the Gamma function as friction terms (US Department of Transportation, 2010).
Because the Gamma functions are rigid mathematical forms, logistic functions are a fairly good replacement for the in-
verse Gamma function (Wen-Liang and Jong-Wuu, 1999). A logistic function is calibrated in this study to demonstrate the
transit trip length’s trend. For instance, a calibrated logistic function with R2 = 93% for the case-study of Winnipeg city
(see below Section 4) is:
1
y¼ ð10Þ
1 þ 20:e10:x
where x = t(s, e)/tmax for (s, e) 2 S is the length (equivalent to travel time in minutes) over the maximum transit trip length
(tmax = 50 for Winnipeg city), and y is the cumulative transit trips over total demand, 0 6 y 6 1. Getting the inverse form
of Eq. (10) and replacing y with a random number 0 6 q 6 1 provides d as follows.
 
q
‘n 1
20q
d¼  tmax ð11Þ
10
The above equation for d aims to seek a wide range of solution space rather than a certain area driven just by using y. The
upper bound of the routes’ length is subject to the condition of the network. Because of excessive operational cost, and
perhaps alternative affordable modes such as walking, the too- short routes are less attractive. Therefore a minimum
route length of 1.8 km is imposed in Eq. (7). The minimum route length is 1.8 km in the existing transit schemes of both
Winnipeg city and Chicago city.

3.2.2. Laying down a route between terminals


By having two terminals to identify a path, the concept of shortest path comes to mind. However, laying down a route on
the shortest path might miss some important transit centers nearby. Because finding shortest paths is a common practice,
these nearby centers can be approached by virtually manipulating links’ travel times to yield access to the nearby transit
centers. For all other purposes the actual travel time is utilized. Thus by finding the shortest path on this manipulated net-
work, the chance of visiting the nearby transit center increases. The following is its formulation.
~tp;q ¼ tp;q
for 0 6 a 6 1; 0 6 b 6 1; where ð12Þ
ðhp;q þ 0:09Þa  ðsp;q þ 1Þb
 ~t p;q is a manipulated travel time on link (p, q) 2 A.
1794 S.A. Bagloee, A. Ceder / Transportation Research Part B 45 (2011) 1787–1804

 For links (p, q) 2 A, an attraction index hp,q is calculated; it depends on how weighty the link’s end points are; that is,

Wp þ Wq
hp;q ¼ ‘ðp;qÞ
ð13Þ
e 0:3
From a shortest path perspective, links that are connected to the transit centers (from weighted end nodes Wp, Wq) are
more attractive (increase of hp,q); however links’ lengths have an inverse effect which nullify the weighted nodes of a
lengthy links. To avoid mathematical degeneration (i.e., zero in denominator), the term (hp,q + 0.09) is used; that is, there
is at least one passenger at end nodes of a link with length of 1 km, see Eq. (13), where 0.09 ffi 1/exp(1/0.3).

 The use of hp,q might deviate routes from passengers’ desired corridors. To avoid such a case sp,q the free-transit-flow is
embedded; it evolved from assigning transit demand over the network (irrespective of any transit system) with tp,q as
links’ travel times. The assignment flows (sp,q) indicate the passengers’ desired path. To avoid mathematical degeneration
the term sp, q + 1 is used; that is, at least one passenger on each link.
 Finally a,b are parameters to regulate impact of both hp,q, sp,q. (In the case study of Winnipeg city in Section 4
below, a = b = 0.5 for equal average of h p,q + 0.09 and sp,q + 1).

In this process, the combination of terminals, path and candidate stops on the manipulated shortest-paths of the candi-
date routes are determined. This process continues up to a certain level (see termination condition in Section 3.2.4), during
which the hierarchy types of the candidate routes are also defined. However there are some more considerations to add be-
fore the generated routes become qualified as a member of R; this is described in the next section.

3.2.3. Determining hierarchy of a route


Initially a greedy-qualification method for the newly generated route (of iteration i) is established. With a travel demand
(at the candidate-stop level, Eqs. (5) and (6)) computed as the amount of directly served passengers (DSPi) and using a
threshold ðDSP i Þ, inadequate candidate routes can be discarded ðDSP i 6 DSPi Þ:DSPi is a decremented threshold for DSPi
and varies by iterations:

DSPiþ1 :¼ DSP i  e DSP iþ1 > 0; e > 0; where ð14Þ


i
DSP ; e are parameters. This is a greedy approach to generate more efficient routes. In early iterations, heavy loaded routes
are generated. As the algorithm proceeds (declining DSP i Þ less loaded routes are generated. The passenger load on the routes
is one of the determinant factors of hierarchy of routes (defined as mass, feeder and local). Starting from iteration i = 1 to
iteration i1 a number of direct and heavy loaded routes, called mass routes, create the skeleton of the transit-route layout.
This moves into other hierarchies as follows.

i 2 ½1; i1  where DSP feeder < DSP i 6 DSP mass ð15Þ


i 2 ði1 ; i2  where DSP local < DSPi 6 DSPfeeder ð16Þ
i local
i 2 ði2 ; i3  where 0 6 DSP 6 DSP ð17Þ
mass feeder
For the case study of Winnipeg city (Section 4 below) conservatively it is assumed that: DSP ¼ 1; DSP ¼ 60;
DSPlocal ¼ 10 and e = 0.05. Thus the algorithm finds plenty of space to generate candidate routes. Consequently i1, i2,
i3(1 < i1 < i2 < i3) are identified as well.
It is worth noting that the way the hierarchy is defined is based on assuming desirable features of the structure of the
network such as the use of corridors, and the size and characteristics of the vehicle fleet. For instance, it is assumed that pas-
senger load of regular buses on heavy-demand routes (mass routes) is at the level of full bus capacity ðDSP feeder ¼ 60Þ; also
half of minibus capacity is assumed as a demarcation threshold between local and feeder routes ðDSPlocal ¼ 10Þ. Because the
definition of DSPi is not capacity restricted it is expected to have a load higher than the capacity (DSPi > 60); this is why the
upper bound is set to infinity ðDSP mass ¼ 1Þ.

3.2.3.1. Generating mass routes. With respect to Eq. (7), terminals are specified as follows.
_
Ms ¼ W s þ 0:03 ð18Þ
_
Me ¼ W e þ 0:03; where ð19Þ

 Ws, We are the nodes’ weights for the respective terminals. The main (mass) routes are usually designed to run between
two major transit centers. There may be several routes in which the load on one direction is high and empty in the reverse
direction (depends on peak periods). To capture these routes a minimum weight is assumed i.e., at least one passenger is
at the terminal see Eqs. (5) and (6), (0.03 ffi 1/exp (1/0.3)). This may also capture routes with low loads at the terminals
and heavy loads in between.
 The friction is defined based on travel time between terminals:
S.A. Bagloee, A. Ceder / Transportation Research Part B 45 (2011) 1787–1804 1795

f ðs; eÞ ¼ ~tðs; eÞpðiÞ ; where ð20Þ


pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
~tðs; eÞ is a geometric mean of terminals’ travel time ~tðs; eÞ ¼ tðs; eÞ:tðe; sÞ, (geometric mean, rather than average, provides
extra incentive to 2-way route with a shorter length in one direction); p(i) is a gradual incremental variable (0.5 < p(i) 6 1) at
iteration i which has a delicate implication on the quality of candidate routes. In early iterations the route-generation rou-
tine is relatively ignorant about the candidate-route length because those routes are aimed to operate as mass transit routes
no matter how lengthy they are. As the routine progresses the route’s length follows the formula:
pðiÞ ¼ 0:5:ð1 þ i=i1 Þ ð21Þ

3.2.3.2. Generating feeder routes. Feeder routes are aimed to connect local areas (or routes) to the mass routes. Concerning the
gravity index (Eq. (7)) terminals are defined as follows.
_ i
Ms ¼ W s  qðas Þ þ 0:03 ðrepresents local stopsÞ ð22Þ

_
Me ¼ ðv ole þ 5:92  aie Þ  ðaie Þq þ 0:03 ðrepresents mass route stopsÞ; where ð23Þ
P P
 vole is free transit passing-through volumes at nodes s, e 2 V under free-transit-flow ðv ole ¼ ðp;eÞ2A sp;q þ ðe;qÞ2A sp;q Þ.
 ais ; aie indicate the number of accesses/egresses provided to the corresponding terminals up to iteration i. For instance, at
iteration i with only two generated candidate-routes: A ? B ? C—layover—C ? B ? A and B ? D—layover—D ? B, we
have aiB ¼ 6 (provided accesses/egresses are: A ? B; B ? C; C ? B; B ? A as well as B ? D; D ? B). Because aie is next
to vole in Eq. (23), it is assumed that each access is equivalent to 5.92 passengers which is the average number of boarding
in the case study of Winnipeg city (Section 4 below).
 Local stops with a few accesses (ais ! 0Þ and more passengers (Ws) are more likely to be selected as a terminal; in the con-
nection to the mass routes (represented by vole), the higher aie and vole, the higher the likelihood to be selected as a terminal.
 Again, embedding the random number q enforces exploring a wide area of the solution space.
 The friction terms of the gravity index is defined as follows:
~
f ðs; eÞ ¼ etðs;eÞ ð24Þ
Exponential perception of the lengths makes the routes more conservative as opposed to the mass routes (feeder routes
should not be as lengthy as mass routes).
The rest of the terms are set as previously described.

3.2.3.3. Generating local routes. Local routes traverse through local and side areas to connect them to the transit network.
Thus, there is less concern about the route directness and load factors than compared to other routes, and the intention
is to cover a wider area. Therefore with reference to Eq. (12) the term of free-transit-flow spq is omitted i.e., b = 0. Thus,
the path using the manipulated shortest path is skewed to the transit centers aiming at collecting more demand according
to Eq. (7); this is formulated as:
_
Ms ¼ W s þ 0:03 for ais < amin ðrepresents local unconnected stopsÞ ð25Þ

_
Me ¼ v ole þ 5:92  aie þ 0:03 for aie > amin ðrepresents connected stopsÞ; where ð26Þ

 Omitting q makes the selection of terminals more straightforward, however the gravity index is still at the stochastic sta-
tus (q in Eq. (7)) to cover a wide range of choices.
 Terminals are separated into already-connected and not-connected terminals by a pre-specified minimum number of
accesses/egresses amin. The termination condition is embedded in this repeated procedure until providing amin number
of accesses/egresses with each candidate-stop (see Section 3.2.4).

3.2.4. Termination condition


At the end of each iteration 1 6 i 6 i3, with hierarchy-related constraints satisfied (Eqs. (14)–(17)) each route hs, ei includ-
ing the maximum gravity-index is qualified as a member of R; that is:
hs; ei ¼ max fGIs;e jEqs: ð14—17Þ holds and trhs; ei 6 1:57  ðtðs; eÞ þ tðe; sÞÞg ð27Þ
8ðs;eÞ2S

where trhs, ei is travel time on the 2-way route hs, ei approximately of the actual shortest-path value. This limitation does not
allow for compromising the routes’ directness by exploring more areas. Using the deviation of a complete circle trip from a
truly direct path (half a circle versus diameter), the allowable deviation is specified as 1.57; that is,

1:57 ¼ ðp:d=2Þ=d; where \d" is the circle’s diameter


1796 S.A. Bagloee, A. Ceder / Transportation Research Part B 45 (2011) 1787–1804

The route generation procedure is repeated until for every candidate-stop there exists 8s 2 S; ais > amin . For the case study
of Winnipeg city, in Section 4 below, it is assumed to have amin = 5.92, implying that there is at least one candidate-route for
every passengers at each stop, because 5.92 is an average number of boarding. Thus the result would be R.

3.3. Route Selection and Frequency Setting (RSFS)

There are numerous candidate routes and a limited budget to accommodate a transit fleet consisting of several vehicle
classes e.g., minibus, bus, metro. Each class is represented by g 2 X and has its own characteristics including total capacity
(1(g)). Without the loss of generality it is assumed, for simplicity, that instead of a limited monetary budget the constraint is
in terms of the number of seats denoted by B. Similarly the cost/value of each candidate-route is defined as the required
capacity satisfying the passenger demand.
The Route Selection and Frequency Setting (RSFS) algorithm is an iterative process partitioned into two phases. The entire
process starts from iteration i = 1 and ends ati = N. In Phase zero, up to iteration N0 < N, the algorithm attempts to produce a
set of stochastic and comprehensive generations to encompass the realm of the solution space. In Phase one (N0 < i 6 N)
based on Phase zero, a hybrid-search-engine, dominated by the AC’s concept of searching for food, is run over the candidate
routes and selects good routes. Contribution of GA during the search is restricted to exert the mutation concept to diversify
the generated scenarios rather than getting bogged down in the local optimums. Notations N and N0 < N) are parameters. At
each iteration (1 6 i 6 N) scenario Si is coded as a chromosome indicating which of the candidate routes is selected (1: se-
lected as genes, 0: otherwise).
Before launching the RSFS module, at each iteration 1 6 i 6 N, an updated capacity for 8r 2 R is required ðci1
r Þ. In the case
study of Winnipeg city (Section 4) an initial value of c0r ¼ 400 is assumed; this is the average number of boarding passengers
in the existing scenario.

3.3.1. Transit assignment and post analysis


There are two general approaches for transit assignment as a means for evaluating route-scenario: capacity-free and
capacity-constrained. For optimum design the routes must be laid down with taking into account passenger demand and
trends. The use of capacity-free assignment makes sense because Phase 0 aims at identifying the capabilities of the routes
through the total number of potential boarding without any restriction. However in Phase 1 an analysis of the actual oper-
ation of the scenarios is taking place and thus requires implementation of a capacity-constrained assignment. EMME3 has
developed a reliable module for both approaches based on optimal strategy (INRO, 2007; Cepeda et al., 2006) which is uti-
lized in this study. All the assignments are carried out on the actual network and not on an imaginary network as utilized in
Eq. (12).
At the end of each iteration (1 6 i 6 N) the following indices are updated:
Point: more points of candidate routes imply more chances for a better scenario (or chromosomes) with respect to the
i
objective function. The pir point of candidate-route r with passenger load of br in iteration i is:
8 i
>
< Ui  Pbr for r 2 Si
bir
Pir ¼ r note : P 0r ¼ 0 ð28Þ
>
: i
Pri1 for r R S

where Ui is total saved travel time (see Eq. (2)).


Required capacity (or cost): cir for "r 2 Si is updated:
8
Si
< 1ci1
r þð1=iÞc r
for r 2 Si
cir ¼ 1þð1=iÞ
note : c0r ¼ 400 ð29Þ
:
ci1
r for r R Si
i
where cSr is the required capacity for r 2 Si at the scenario of iteration i. The process continues (i ? N) while 1/i makes the
i
final required capacities ðcir Þ less sensitive to csr subject to the conditions of scenario Si. This helps to have a convergent pro-
cess within the span of the pre-specified number of iterations (N):
qffiffiffiffiffiffiffiffiffiffiffiffiffiffi
i
cSr ¼ l^f ir  lf ir  fri;ðgÞ  1ðgÞ for r 2 Si ð30Þ

i;ðgÞ
where l^f ir ; lf ir and fr are maximum load-factor, average load-factor and required fleet size of class g 2 X for route r 2 Si. Be-
cause 0 6 lf r 6 lf r < 1, the geometric mean of l^f ir ; lf ir is used.
i ^i

3.3.1.1. Multiclass design. Both multiclass design and frequency setting are embedded in this analysis simultaneously. At the
ðgÞ
end of each iteration (1 6 i 6 N) the required headway of r 2 Si concerning every available transit class ðhdwyr Þ is computed:

ðgÞ 60
hdwyr ¼  i for r 2 Si ; g2X ð31Þ
cr þ 1 =1ðgÞ
S.A. Bagloee, A. Ceder / Transportation Research Part B 45 (2011) 1787–1804 1797

To avoid mathematical-degeneration the capacity is assumed as cir þ 1. Then for "r 2 Si, "g 2 X the amount of unused
capacity is calculated. The class with the minimum unused capacity is allocated to the corresponding candidate-route, that
is,
8  
>
< g so that minfjci  1ðlÞ  max 1;
r
60
ðgÞ jg for r 2 Si
v ehir ¼ > 8g2X hdwyr Note : v eh0r ¼ arbitrary g 2 X ð32Þ
: i1 i
v ehr for r R S
i
where v ehr is the allocated vehicle class for r 2 Si.
It is worth noting that there are additional factors which affect the headways; that is, fleet speed, route’s length and lay-
over time. These factors have been taken into account during the assignment process and have been added to the load of the
i
route ðcSr Þ then cir of Eq. (13). In other words, more passengers are attracted to routes with shorter travel time (higher
speeds). For the case of Winnipeg the layover time varies from 5 to 40 min and the cruising speed is enforced by transit travel
time functions (see INRO (2007)).
Vehicle allocation is not restricted in this analysis to the size/type of vehicle. For instance, one may assume the allocation
of minibuses or vans only for short routes. Such strategies can be implemented within the proposed analysis.

3.3.2. Phase 0: Generating initial stochastic chromosomes


This phase attempts to obtain a good knowledge of the capabilities of the routes using their collective points from differ-
ent chromosomes. In a chromosome laden with many routes the points would become ambiguous. Thus the budget ( B b i total
available capacity) is confined to around 25% (not less than 10% of the available budget (B); randomly speaking (i.e.,
q2 25%):
b i ¼ ½maxðq2 ; 0:1Þ
B
B ð33Þ
At each iteration (1 6 i 6 N0) route candidates are selected as follows.

 Step 0: Compute confined budget B b i and initialize a new and empty chromosome Si,
 Step 1: Take a candidate-route (r 2 R; which has not been taken) randomly,
 Step 2: Forr 2 R, if required capacity is not less than the budget ðcir < B b i Þ go to Step 4,
 Step 3: Put r into the chromosome (r ? Si) and update the budget ( B b i :¼ B b i  ci Þ,
r
 Step 4: If still some of the candidate routes have not been taken go to Step 1,
 Step 5: End.

During the current iteration (i) at the end of this algorithm a temporary scenario (Si) has been generated, pending check-
ing whether or not it has been generated in the past iteration. If the scenario has not been generated previously then, Si is
retained in the database of generated scenarios. Otherwise the following action is taken.
Resume/Mutation: If scenario Si has been generated in past iterations the algorithm for the current iteration is resumed
or, by a narrow chance (10%), the current chromosome is mutated. Mutation is a key component of GA and aims at prevent-
ing stagnation at the local optima, and exploring the extended solution space. The designed mutation module, of this anal-
ysis, simply and randomly takes out a selected candidate-route from the chromosome (Si) and brings an unselected
candidate-route into Si with respect to the budget.
For a given scenario, the capacity-free assignment procedure and post analysis are conducted until all the required indices
are updated (Eqs. (28)–(32)). This process is repeated until each r 2 R contributes at least to one Si,1 6 i 6 N0. Thus, after
reaping enough provision of chromosomes (for the case study of Winnipeg city it becomes N0 = 15), the algorithm proceeds
to Phase1 to seek out a good solution.

3.3.3. Phase 1: Seeking good chromosomes


According to AC, each ant, before being restricted for food, evaluates expected food at different sources ahead. Sources
here are interpreted as candidate routes, and for evaluation purposes an index called collective points denoted by CP ir is de-
fined. In order to utilize the memory of previous scenarios’ performance, as a proxy of pheromones, CP ir for 8r 2 R at iteration
N0 < i 6 N is computed by using the top six chromosomes stochastically combined and mixed with the required capacity.
Adopting only the top six chromosomes in the CP ir arises from the following notion: ideally it is assumed that every candi-
date stop is provided with amin access/egress; each access/egress, at most, needs one route which is supposed to be of top
chromosomes. For the Winnipeg case study amin = 5.92 6 (see Section 3.2.4). The CP ir is as follows:

X
6
CPir ¼ ðq  Pjr  ðUi;j =Ui;1 Þ2 Þ ð34Þ
j¼1

where Ui,1 > Ui,2 >. . .> Ui,6 are the amount of the objective functions for the six top chromosomes for iteration i. Adjust-
ment with the square of the proportions (Ui,j/Ui,1)2 aims to give more weights to the routes engaged with stronger
chromosomes.
1798 S.A. Bagloee, A. Ceder / Transportation Research Part B 45 (2011) 1787–1804

During the selection process of candidate routes, an operational limitation of headways is considered with a policy-head-
way of 60 min. There is no bound on the minimum headway because for high-demand routes more parallel routes can be
designed; thus,

ðv ehri1 Þ
hdwyr 6 60=ð2qÞ1=i ð35Þ

At early iterations (higher values of 1/i) the algorithm staggers between different classes; therefore the headway is unpre-
dictable. When i ? N the effect of the random term (2q)1/i diminishes (note that 2q 1).
The candidate routes are then sorted by their collective points in a descending order. The algorithm starts from the begin-
ning of the sorted list. Each route complying with the policy-headway and available budget is chosen in the current scenario
until the entire list has been checked.
The situation of scenario duplication is being treated by the Resume/Mutation procedure, described above. To avoid
degeneration, this procedure is repeated up to a certain running time, say 5½R. At this stage if no scenario has been gener-
ated the algorithm terminates. However the chance of this happening for real transit networks is nil (for instance, the solu-
tion space of the Winnipeg case study with 245 candidate routes is 2245 = 5.6E + 73, thus the chance of getting the same
scenario by exerting the mutation module of this ample space is almost zero).
From a practical and operational perspective, a capacity-constrained assignment procedure is implemented
(Cepeda et al., 2006) using the generated scenario (set of routes); then the required indices are updated as described
in Section 3.3.1.
This phase for the case study of Winnipeg city was repeated N  N0 = 25 times until the results showed no further
improvement. The ‘‘no more improvement’’ criteria is used if after 2amin iterations a better solution can’t be reached.

4. Case studies

The main purpose of this study is to find an efficient and feasible solution for real-life transit networks. Consequently the
developed algorithm is tested on a case study of Winnipeg city. In addition the developed methodology is compared with
previous studies using the transit network of Mandl (1980) as a benchmark. We also examined the generality of the meth-
odology in the broader context of optimizing the existing rail system of the greater city of Chicago in which the recent real
dataset (pertaining to the year 2010) of the city was utilized.
The algorithms developed in this work are in Visual Basic 6.0 (VB) and consist of three modules (Stop-Positioning, Route-
Generation and RSFS); each module comprises several sub-modules and procedures. The VB is linked to a worksheet in MS-
Excel to obtain input data easily. In addition the generated routes are saved and stored by separated MS-Excel files. The VB is
in dynamic interaction with a MS-ACCESS file called ‘‘transit.mdb’’ which is used as a database (to manage vast amount of
data and computations especially for the Stop-Positioning and Shortest-Path analyses). The VB is integrated to EMME3 for
coding and analyzing the scenarios, determining the shortest-path and conducting the transit-assignment process. This
set-up has two primary advantages. First, the developed package provides a user-friendly interface for ordinary users. Sec-
ond, the sequential process and accessible input and output of each module provides a valued tool for the user to alter or
modify the process, e.g., to insert routes and/or stops into the final scenario. For this software package a PC with 1.86GHz
CPU and 2.00GB of RAM is used.

4.1. Winnipeg case study

The data of Winnipeg, the capital and largest city of Manitoba, Canada, pertaining to the year 1976 is standard for EMME3
(INRO, 2007) and is utilized in this study. Winnipeg city was chosen and elaborated as a case study because of its relatively large
size and availability of data to researchers and practitioners. In addition the data of Winnipeg road network is available on the
Internet (Bar-Gera, 2011). The case consists of 154 zones, 903 nodes and 2975 links as well as 5394 nonzero transit-OD pairs
with 18,210 total transit demands. The existing transit fleet includes 67 of 2-way routes (1582 km) with two different transit
vehicles, bus and minibus, with capacity of 1(1) = 60 and 1(2) = 37, respectively. The fleet is comprised of 435 buses and 35 mini-
buses. The current set of stops was taken as the initial stops to allow for the comparison of the solution with the existing system;
the algorithm selected 256 stops out of 348 existing stops. The computational process took about 40 h. It was assumed 60% as
average load-factor of the existing fleet with the budget set as B = 16,160 = (435
60 + 35
37)60%. For this case study it is as-
sumed that for Eq. (1) x1 = x2 = x3 = 2. Other parameters have been introduced above. All the other factors such as transit-
vehicle extra delays, dwell times and prohibited turns were kept the same as the existing Winnipeg scenario (more details
in INRO, 2007).
The algorithm generated 245 candidate routes (3838 km) in about 70 min. Finally the RSFS module was executed for
about 29 h. Fig. 2 illustrates the progress of the algorithm over successive iterations. Fig. 2a shows the total amount of saved
time (user-min) for the case without any transit system subject to capacity-free assignment; the average of saved time be-
came 1,864,583, implying, in the absent of any transit system, that each person bears 102 min walking (102 = 1,864,583/
18,210). In addition Fig. 2a demonstrates an upward trend of the RSFS algorithm. However, the rapid convergence of Phase
1 does not guarantee the optimal solution. It might either indicate stagnation over a good solution or efficiency of the
S.A. Bagloee, A. Ceder / Transportation Research Part B 45 (2011) 1787–1804 1799

(a) 2

Saved traveltime
(1.0e6 user-min)
1.8
1.6
1.4
1.2
1
0.8
1 4 7 10 13 16 19 22 25 28 31 34 37 40
phase 0 phase 1
iteration (i)

(b) 2.36
Existing situation
Total travel time
(1.0e6 user-min)

2.405382
2.25
14%

2.14

2.03
16 18 20 22 24 26 28 30 32 34 36 38 40
iteration (i)
Fig. 2. Winnipeg case-study features of the methodology developed (a) saved travel time, by progressive iterations, between do-nothing and the successive
iterations of Phases 0 and 1 for capacity-free analysis (b) difference of total travel time between the generated scenarios and the existing route network, by
progressive iterations of Phase 1 using the capacity-constrained assignment.

algorithm to yield a good solution. Fig. 2b depicts the variations of actual total travel time (i.e. subject to capacity constraint
assignment) in Phase 1 compared with the existing situation. The best solution (at iteration 26) improves the performance of
the existing transit routes by 14%. This best solution includes 298 buses and 218 minibuses operating on 106 routes, mean-
ing a 5.23% reduction of the existing fleet. From the hierarchy perspective, out of 106 routes, 14 are mass routes, 30 are fee-
der routes and the rest are local routes.
The results of a 14% reduction in user travel time and 5.23% reduction of the fleet size were attained with 9.78% additional
crew (bus/minibus drivers); this implies a tradeoff between the user and operator cost. This tradeoff can be investigated in a
future research.

4.2. Mandl benchmark

Mandl’s benchmark is a small and dense network of 15 nodes and 15,570 total demands which has been widely cited by
researchers. Because of the similarity of approaches (route-generation and route-selection) the proposed methodology is
compared to the studies of Shih et al. (1998), Baaj and Mahmassani (1991) and Mandl (1980). Table 1 shows the results
and comparison in which the developed procedures result in a transit network of 12 routes.
The comparison of Table 1 shows that the proposed methodology, with almost the same fleet size as the others, provides a
moderately better solution than others particularly in terms of lesser transfer time. This study proposes more routes than
other studies in the process of seeking more direct routes. It is also interesting to observe that the proposed solution yields
the highest number of double-transfers apparently captured in the out-of-vehicle (transfer) time. The relatively high penalty
cost of transfers interpreted by Eqs. (9)–(11) provides the solution with more direct routes which makes the transit network
more attractive to the users.
A similar result in terms of more routes has been observed in the Winnipeg case study. General and primary reason can be
attributed to the adopted approach of laying down the routes as close to the transit centers and passengers desirable paths
(Eq. (7)) as possible.

4.3. Chicago’s 2010 rail system

According to the US census data, Chicago is the largest city in the US state of Illinois. With nearly 2.7 million residents in
the year 2010, it is the third most populous city in the USA, and the 27th most populous urban agglomeration in the world.
The city benefits from a very advanced and extended transit system with, on an average weekday, 1.6 million riders. The
Chicago transportation model consists of 1961 zones, 13,487 nodes and 52,742 links as well as 142,041 nonzero
1800 S.A. Bagloee, A. Ceder / Transportation Research Part B 45 (2011) 1787–1804

Table 1
Comparison with other studies using Mandl benchmark.

% of Demand Total time elements (min)


Solution Fleet No. of Zero-transfer One-transfer Two-transfer Travel In-vehicle Out-of-vehicle Transfer
size lines time time time (penalty)
Shih et al. [9]a Coordinated 87 6 82.59 17.41 0 225,102 191,826 19,726 13,550
Shih et al. [9]a Uncoordinated 84 6 82.59 17.41 0 203,936 170,328 20,058 13,550
Baaj et al. [14] 82 7 80.99 19.01 0 217,954 180,356 22,804 14,800
Mandl [33] 99 4 69.94 29.93 0.13 219,094 177,400 18,194 23,500
This studyb 87 12 83.66 15.21 0.95 202,255 167,198 24,591 10,465
a
Coordination refers to changing departure times for minimizing transfer times.
b
The proposed solution of 12 routes comprised of:

Route 1: fleet of 3 vehicles >> 9–7–5–2–1–0;


Route 2: fleet of 2 vehicles >> 9–6–14–8;
Route 3: fleet of 2 vehicles >> 6–14–5–2–1;
Route 4: fleet of 5 vehicles >> 6–14–5–3–4;
Route 5: fleet of 10 vehicles >> 0–1–2–5-7–9–13;
Route 6: fleet of 21 vehicles >> 12–9–7–5–2–1–0;
Route 7: fleet of 7 vehicles >> 11–10–9–6;
Route 8: fleet of 5 vehicles >> 10–9–7–5–2–1–0;
Route 9: fleet of 8 vehicles >> 8–14–6–9–12;
Route 10: fleet of 5 vehicles >> 3–5–7–9–10;
Route 11: fleet of 13 vehicles >> 10–9–7–5–3–4;
Route 12: fleet of 6 vehicles >> 9–10–11;

transit-OD pairs with 134,390 total (hourly) transit demands. The transit system comprises 648 stops and 508 routes, of
which 172 routes delineate the rail system (note: routes pertaining to the same terminals and the same path but different
stop plans have been coded separately). Fig. 3 depicts the existing transit route scheme and size of the road network. In the
model the two rail classes are known as CTA-Rail and Metra-Rail with capacities of 125 and 150 respectively. The fleet size of
the rail system (CTA-Rail and Metra-Rail collectively) is 1118 vehicle. The existing scheme distinguishing these two classes
has been depicted in Fig. 4a and b.
We explore the generality of the proposed methodology to seek a better and efficient rail system on the scale of the
sizable case of Chicago. Conservatively, in order to run a worst case scenario, we keep the non-rail system unchanged
and try to achieve the same operational level of the existing scenario through optimizing the rail system. More precisely,
an efficient rail system with a smaller required fleet is sought. First we run the clustering module that yielded 330 can-
didate-stops out of 648 existing stops. According to the established methodology, the algorithm consists of two parts:
Part 1 – Generating candidate routes and Part 2 – selecting the best sub set of the candidate routes. Because the rail
routes can be categorized as mass routes due to the ridership and routes’ length, we run Part 1 only for mass routes
(i.e. DSPmass ¼ 1Þ and similar to the Winnipeg case, the mass route threshold is assumed as Metra-Rail’s capacity
(DSPfeeder ¼ 150, see Eqs. (14)–(17)). Therefore 70 mass routes were generated which took the process around 16 h. How-
ever the prior preparation to compute the transit demand at stop-level for the sizeable data set of 330 candidate stops
m;n
was time consuming (almost 10 days). Fortunately computing dc ðbÞ in Eq. (5) is independent of other pair candidate-
stops (i.e. m,n 2 S) and that makes parallel computation possible. Thus we employed 5 computers and the computation
was carried out in 2 days. At the end of Part1 we mixed the generated candidate-routes (70 routes) with the existing
routes (172 routes) and proceeded to the next part (This was simply made possible by changing interpretation of super-
script 0 in Eq. (2) to ‘‘Existing-non-Rail’’ scenario).This mixing contains two key points. First, the existing routes them-
selves are eligible to be considered as at-least candidate routes since they stem from an engineering process and may
cater to some specific considerations of the authorities. Second, the contribution share of the algorithm generated-route
in the final scheme can be regarded as an evaluation index for the proposed algorithm. This contribution share in the set
of candidate-routes is 30% (i.e. 30% = 70/(70 + 172)%). In Part 2 of the algorithm, since the size of the set of candidate-
routes is comparable to the Winnipeg case (242 = 70 + 172 versus 245 in Winnipeg) we adopted the same number of
iterations: N0 = 15 and N = 40. Assuming that up to 80% of the existing fleet is available; Part 2 of the algorithm was
launched. The process was terminated in 5 h and the best solution was found at iteration 29. Fig. 4c and d depicts
the identified rail route scheme which consists of 54 routes requiring a fleet size of 882 vehicles (21% reduction) while
the total user-time has been kept almost identical (i.e. 21,622,146 passenger-minutes in the proposed scheme versus
21,527,482 in the existing scheme). It worth noting that ‘‘passenger-minutes’’ is the total imposed general times or
charges including walking and in-vehicle times as well as transfer (or ticket) charges (see Eq. (1), with the same assump-
tions as in the Winnipeg case: x1 = x2 = x3 = 2). The proposed scheme inherited only 13 routes of the existing scheme’s
S.A. Bagloee, A. Ceder / Transportation Research Part B 45 (2011) 1787–1804 1801

Fig. 3. The extend of the existing Chicago transit system and the road network.

routes (24%). It is important to note that the achievement of 21% reduction in this case and 14% user-time improvement
in the Winnipeg case should not be taken out of the actual context. Relevant authorities are always required to consider
other issues such as ‘‘Equity’’, ‘‘Sustainability’’ and so on, when it comes to transit planning. These issues make the re-
search open to further investigation. Nonetheless, the proposed methodology can accommodate some special constraints
such as enforcing some routes at any cost as we did with non-rail routes or deliberately increasing the chance of some
specific candidate-routes via various tools such as mass and points. At least from a purely transportation point of view,
the case-studies undertaken indicate that the proposed methodology is capable of handling sizable and realistic cases
which was the main target of this study.
1802 S.A. Bagloee, A. Ceder / Transportation Research Part B 45 (2011) 1787–1804

Fig. 4. Chicago transit system (a) existing rail routes, rail class of CTA-rail (b) existing rail routes, rail class of metra-rail (c) algorithm proposed rail scheme,
rail class of CTA-rail (d) algorithm proposed rail scheme, rail class of metra-rail.

5. Conclusion

A heuristic methodology for a complex problem of transit-network design is developed for handling actual-size road net-
works. This study tries to address the main concerns in the transit planning (such as routes hierarchy, multiclass planning
and so on) which have not been considered collectively in the past studies. A systematic transit-planning module consists
and synchronized of various components and concepts such as clustering, Newton gravity and Genetic Algorithm is devel-
oped. The methodology proposed takes into account the major concerns of the transit authorities such as budget constraints,
level-of-service standards and the attractiveness of the transit routes. In addition this approach considers other important
aspects of the problem including categorization of stops, multiclass of transit vehicles, hierarchy planning, system capacity
(which has been largely ignored in past studies) and the integration between route-design and frequency-setting and anal-
yses collectively.
In the developed methodology there are two main parts. In the first part the concept of clustering is utilized to construct a
good scheme of stops, and an algorithm is built comprised of two components: route-generation and route-selection. The
concept of Newton gravity is used to determine the terminals (end points of a route). The concept of a manipulated short-
est-path between terminals is introduced to favor skewed routes that cross transit centers.
In the second part of the methodology a search algorithm is launched over the generated routes to attain a good subset of
routes within the budgetary constraints. The overall approach, implemented by software, was tested on a case study using
the actual-size transit network of the city of Winnipeg, Canada. The results show that under same conditions (budget and
S.A. Bagloee, A. Ceder / Transportation Research Part B 45 (2011) 1787–1804 1803

constraints) the proposed set of routes (scenario) resulted in a reduction of 14% total travel time when compared to the
existing transit network. In addition it resulted in a 5.23% reduction of the fleet size. This reduction was however accompa-
nied by the need for a 9.78% increase in crew, implying a tradeoff between the user and operator cost. The latter can be inves-
tigated in a future research.
The developed approach is also compared with other studies using the Mandl (1980) benchmark. This comparison
shows that the proposed methodology, with almost the same fleet size as the others, provides a better solution
than others especially in terms of more direct routes; however a marginal part of the demand (0.95%) faced double
transfer.
Finally, we examined the generality of the methodology in the broader context of optimizing the existing rail system
of the greater city of Chicago in which its recent real dataset was utilized. The algorithm generated-routes mixed with
the existing routes eventually evolved to an efficient scheme in which, with a smaller fleet (21% reduction), the
same transportation standard level expected from the existing scheme was obtained. Regardless of other considerations
such as ‘‘Equity’’ and ‘‘Sustainability’’, at least from a purely transportation perspective, the numerical results indicate
that the proposed methodology is capable of handling sizable and realistic cases, which was the main objective of this
study.

Acknowledgments

The Authors are indebted to Dr. Lorna Richardson Director of Sustainability, Middle East at Parsons and Dr. Ehsan Maz-
loumi at Monash University, Australia, whose comments improved the paper’s presentation. The authors would like to thank
the anonymous reviewers and the editor for their insightful suggestions that raised the motivation to apply the algorithm to
the case of Chicago. This would not be possible without the great assistance that the authors received from Chicago Metro-
politan Agency for Planning (CMAP) especially Claire Bozic.

References

Baaj, M.H., Mahmassani, H.S., 1991. An AI based approach for transit route system planning and design. Journal of Advanced Transportation 25 (2), 187–210.
Baaj, M.H., Mahmassani, H.S., 1995. Hybrid route generation heuristic algorithm for the design of transit networks. Transportation Research Part C 3 (1), 31–
50.
Bar-Gera, H., 2011. Transportation Network Test Problems. <http://www.bgu.ac.il/bargera/tntp> (accessed 11).
Bielli, M., Caramia, M., Carotenuto, P., 2002. Genetic algorithms in bus network optimization. Transportation Research Part C 10 (1), 19–34.
Carrese, S., Gori, S., 2002. An urban bus network design procedure. Applied Optimization 64, 177–196.
Ceder, A., Gonzalez, O., Gonzalez, H., 2002. Design of bus routes-methodology and the Santo Domingo Case. Transportation Research Record 1791. Journal of
the Transportation Research Board, 35–43.
Ceder, A., 2007. Public Transit Planning and Operation: Theory, Modeling and Practice. Elsevier, Butterworth-Heinemann, Oxford, UK.
Ceder, A., Wilson, N.H.M., 1986. Bus network design. Transportation Research Part B 20 (4), 331–344.
Cepeda, M., Cominetti, R., Florian, M., 2006. A frequency-based assignment model for congested transit networks with strict capacity constraints:
characterization and computation of equilibria. Transportation Research Part B 40 (6), 437–459.
EMME3 User’s Guide, Revision 1.22. INRO, 2007. Montreal, Quebec, Canada.
Fan, W., Machemehl, R., 2006b. Using a simulated annealing algorithm to solve the transit route network design problem. Journal of Transportation
Engineering 132 (2), 122–132.
Furth, P., Rahbee, A., 2000. Optimal bus stop spacing through dynamic programming and geographic modeling. In: Transportation Research Record, Journal
of the Transportation Research Board, No. 1731, Transportation Research Board of The National Academies, pp. 15–22.
Gao, Z., Sun, H., Shan, L., 2003. A continuous equilibrium network design model and algorithm for transit systems. Transportation Research Part B 38 (3),
235–250.
Guan, J.F., Yang, H., Wirasinghe, S.C., 2003. Simultaneous optimization of transit line configuration and passenger line assignment. Transportation Research
Part B 40 (10), 885–902.
Guihaire, V., Hao, J.K., 2008. Transit network design and scheduling: a global review. Transportation Research Part A 42 (10), 1251–1273.
Hasselström, D., 1981. Public Transportation Planning – A Mathematical Programming Approach. Ph.D. Thesis, University of Göteborg, Sweden.
Knoppers, P., Muller, T., 1995. Optimized transfer opportunities in public transport. Transportation Science 29 (1), 101–105.
Magnanti, T.L., Wong, R.T., 1984. Network design and transportation planning: models and algorithms. Transportation Science 18 (1), 1–55.
Mandl, C.E., 1980. Evaluation and optimization of urban public transportation networks. European Journal of Operational Research 5 (6), 396–404.
Murray, A.T., 2003. A coverage model for improving public transit system accessibility and expanding access. Annals of Operations Research 123 (1), 143–
156.
Pattnaik, S.B., Mohan, S., Tom, V.M., 1998. Urban bus transit route network design using genetic algorithm. Journal of Transportation Engineering 124 (4),
368–375.
Rapp, M.H., Gehner, C.D., 1976. Transfer optimization in an interactive graphic system for transit planning. In: Transportation Research Record, Journal of
the Transportation Research Board, No. 619, Transportation Research Board of The National Academies, pp. 27–33.
Saka, A., 2001. Model for determining optimum bus-stop spacing in urban areas. Journal of Transportation Engineering, ASCE 127 (3), 195–199.
Salzborn, F.J.M., 1980. Scheduling bus systems with interchanges. Transportation Science 14 (3), 211–220.
Shih, M., Mahmassani, H.S., Baaj, M., 1998. A planning and design model for transit route networks with coordinated operations. transport. In:
Transportation Research Record, Journal of the Transportation Research Board, No. 1623, Transportation Research Board of The National Academies, pp.
16–23.
Silman, L.A., Barzily, Z., Passy, U., 1974. Planning the route system for urban buses. Computer and Operations Research 1 (2), 201–211.
Tom, V.M., Mohan, S., 2003. Transit route network design using frequency coded genetic algorithm. Journal of Transportation Engineering 129 (2), 186–195.
US Department of Transportation, Travel Model Improvement Program (TMIP). <http://tmip.fhwa.dot.gov/clearinghouse/docs/mvrcm/ch4.stm> (accessed
07.10).
Van Nes, R., Hamerslag, R., Immers, B.H., 1988. Design of public transport networks. In: Transportation Research Record, Journal of the Transportation
Research Board, No. 1202, Transportation Research Board of The National Academies, pp. 74–83.
1804 S.A. Bagloee, A. Ceder / Transportation Research Part B 45 (2011) 1787–1804

Wen-Liang, H., Jong-Wuu, W., 1999. Some properties of the extended generalized logistic-gamma distribution with applications. Information and
Management Sciences 10 (4), 1–17.
Xiong, Y., Schneider, J.B., 1993. Transportation network design using a cumulative algorithm and neural network. In: Transportation Research Record,
Journal of the Transportation Research Board, No. 1364, Transportation Research Board of The National Academies, pp. 37–44.
Yan, S., Chen, H.L., 2002. A scheduling model and a solution algorithm for inter-city bus carriers. Transportation Research Part A 36 (9), 805–825.
Zhao, F., 2006. Large-scale transit network optimization by minimizing user cost and transfers. Journal of Public Transportation 9 (2), 107–129.

You might also like