Transit route design based on trip-based agglomeration of travel

demand

Extended abstract submitted for presentation at the 12th Triennial Symposium on

Transportation Analysis conference (TRISTAN XII)
June 22-27, 2025, Okinawa, Japan

December 27, 2024

Keywords: Transit routes, Trip-based agglomeration, Road networks, Mixed integer linear pro-
gram

1 Introduction
The design of routes in transit networks is an important problem as it affects other aspects of
transit system design such as the setting of timetables, planning of crews, etc. (Vermeir et al.,
2021). Traditionally, this problem has been tackled considering road network data and transit
demand matrix as inputs (Chakroborty & Wivedi, 2002, Vermeir et al., 2021). Note that these
approaches require an origin-destination (OD) matrix representation of transit demand and thus
require the trip ends to be agglomerated towards the a priori decided zone centroids (and then to
the potential bus stops). However, a novel trip-based agglomeration (TBA) method proposed by
Kumar et al. (2024) represents each trip as an ordered pair of origin and destination in the form
of a line segment and these line segments are agglomerated to form corridors (note that each
corridor is a line segment whose length is dependent upon the endpoints of the agglomerated
trips). Kumar et al. (2024) showed that the proposed TBA method can identify major flow-
carrying corridors in a single step as opposed to multi-step post-processing that may arise in the
OD-based agglomeration (ODA) methods.
This paper proposes a novel method to design transit routes given the corridors obtained
from the TBA method. Since corridors are line segments that may not be aligned with the
underlying road network, it is required to determine a route on the underlying road network that
closely represents the given corridor while satisfying the following desirable conditions. One of
the conditions is that the determined route should pass through some of the important points
on the corridor such as the endpoints, points with high-demand activity, etc. Another condition
that the developed route should satisfy is that its length should be as small as possible. While
ensuring the above two conditions, it may happen that there may be many turns along the
determined route, and therefore the last condition is to keep the number of turns in the route
as low as possible. In order to ensure that all these conditions are met as much as possible, an
optimization problem is proposed to design the transit routes that maximizes the weighted sum
of the number of important points on corridor that are visited in the determined route minus
the length of the determined route minus the number of turns in the determined route.
A major difference between the proposed optimization problem and the studies that require
OD matrix representation of transit demand is that the latter studies benchmark their results on
small-scale networks like Mandl’s network with 15 nodes (Vermeir et al., 2021), Mumford network
with 127 nodes (Ahmed et al., 2019), etc., whereas in this work transit routes are designed for a
large-scale network with approximately 20, 000 nodes.
Note that the proposed optimization formulation has high-level analogies to the orienteering
problem (OP) (Vansteenwegen et al., 2011) but there are important differences also as explained

TRISTAN XII Symposium Original abstract submittal

 2

next. One of the versions of OP involves finding a route between two given points that maximizes
the number of points that are visited from a pre-determined set of points under the constraint
that the length of the determined route does not exceed a predetermined threshold. However, the
proposed optimization problem considers the number of turns in the objective function because
the determined route is developed on a given road network where the angle of turns that are
made is important unlike that in OP. Furthermore, the length of the determined route is part
of the objective function in the proposed optimization formulation whereas there is a constraint
on the length of the determined route in OP. In the next section, the developed optimization
formulation is presented.

2 Problem Formulation
This section describes the proposed optimization formulation for developing transit routes. Let c
be a directed transit demand corridor as obtained from the TBA method of Kumar et al. (2024)
(note that a directed corridor is a directed line segment). The source and the target endpoints
of the given corridor c are denoted as o and d, respectively. Let G = (V, E) be a directed graph
representing the underlying road network. For all (i, j) ∈ E, let lij be the distance from node i
to node j. Suppose s ∈ V and t ∈ V represent the nodes closest to the source endpoint and the
target endpoint, respectively, of the given corridor c. The output of the proposed optimization
problem is a route that starts from node s and ends in node t while satisfying the conditions
described in the subsequent discussion. Let P be a given set of important points on corridor c
such as points with a high-demand activity (note that set P does not include the endpoints o
and d). Let Q ⊂ V be the set of nodes that are closest to the points in set P . Let H be the set of
all the pair of edges ((i, j), (j, k)) (such that (i, j) ∈ E, (j, k) ∈ E and i ̸= k) where a turn is said
to be present. Let wp and wt be the scaling factors for different terms in the objective function.
The target of the proposed optimization formulation is to determine a route connecting s to t
so that the sum of the number of nodes from set Q that lie in the determined route minus wp
times the length of the determined route minus wt times the number of turns in the determined
route is maximized. The proposed optimization formulation is given as follows.
" # " # " #
X X X
Maximize ρk − wp xij lij − wt fijk (1)
∀k∈Q ∀(i,j)∈E ∀((i,j),(j,k))∈H

Subject to:

X
xsj = 1 (2)
∀j∈V \{s}
X
xis = 0 (3)
∀i∈V \{s}
X
xit = 1 (4)
∀i∈V \{t}
X
xtj = 0 (5)
∀j∈V \{t}
X X
xik = xkj ∀k ∈ V \{s, t} (6)
∀i∈V \{t} ∀j∈V \{s}
X
ρk ≤ xik ∀k ∈ Q (7)
∀i∈V \{t}

fijk ≥ xij + xjk − 1 ∀((i, j), (j, k)) ∈ H (8)

TRISTAN XII Symposium Original abstract submittal

 3

uj ≥ ui + xij n − n + 1 ∀i, j ∈ V and i ̸= j (9)

ρk ∈ {0, 1} ∀k ∈ Q (10)
xij ∈ {0, 1} ∀(i, j) ∈ E (11)
fijk ∈ {0, 1} ∀((i, j), (j, k)) ∈ H (12)

The expression in Equation 1 presents the objective function as described before. Here, for
all k ∈ Q, ρk is equal to 1 if node k lies in the determined route, otherwise it is equal to 0. Also,
for all (i, j) ∈ E, xij is equal to 1 if the edge (i, j) lies in the determined route, otherwise it is
equal to 0. For all ((i, j), (j, k)) ∈ H, fijk is equal to 1 if the edges (i, j) and (j, k) lie in the
determined route, otherwise it is equal to 0. Thus, the first term from the left is the number
of nodes from set Q that lie in the determined route, the middle term is wp times the length of
the determined route and the last term is wt times the number of turns in the determined route.
Equation 2 ensures that only one edge that lies in the determined route leaves the source node s.
Equation 3 ensures that none of the edges that lie in the determined route enter the source node
s. Equation 4 ensures that only one edge that lies in the determined route enters the source node
t. Equation 5 ensures that none of the edges that lie in the determined route leave the source
node t. Equation 6 ensures that for each node k ∈ V \{s, t}, the number of incoming edges to
node k that lie in the determined route is equal to the number of outgoing edges from node k
in the determined route. The constraint in Equation 7 ensures that for all k ∈ Q, ρk is equal to
0 if there is no incoming edge to node k in the determined route. The constraint in Equation 8
ensures that for all ((i, j), (j, k)) ∈ H, fijk is equal to 1 if both the edges (i, j) and (j, k) lies
in the determined route. Equation 9 is a sub-tour elimination constraint that ensures no cycles
are formed in the determined route. Here, ui , ∀i ∈ V are unrestricted decision variables used for
eliminating subtours in the determined route. Equations 10 to 12 are the definitional constraints
for the decision variables. Since some of the decision variables in the above formulation are
integers (except the variables ui , ∀i ∈ V ), and the objective function and the constraints are
linear, the proposed optimization formulation is a Mixed Integer Linear Program (MILP).

3 Results and discussions

In this section, the proposed optimization formulation is applied on a given set of corridors to
determine the transit routes for a large-scale road network. A set of 40 corridors obtained by the
TBA method of Kumar et al. (2024) for Manhattan, New York City, and the underlying road
network for Manhattan (consisting of approximately 20, 000 nodes and 30, 000 directed edges
(Dataset 1, 2014)) are used as inputs for the proposed optimization formulation. Note that 40
corridors are taken as input as there are about 40 transit routes currently in Manhattan. The
optimization problem is solved1 for each corridor and this process is repeated 40 times to obtain
a set of 40 transit routes. Figure 1a shows the computed transit routes in red color. Note
that the average computation time required to determine each transit route is approximately 75
seconds. Figure 1b shows the existing transit routes in Manhattan (Dataset 2, 2020). It can be
seen that at a high-level, the obtained routes look similar to the current routes but there are
noticeable differences. That is because the determined transit routes are computed solely based
on the information on transit demand and the underlying road network whereas the existing
transit routes are designed based on other information as well such as site-specific constraints
that may not allow transit routes to pass through some specific locations, welfare policies that
may require providing routes at low demand areas, etc. Therefore, the proposed optimization
framework only provides an initial estimate of the transit routes that should be suitably modified
by the transit route planners to consider various practical constraints and government policies.
Note that the obtained transit routes are statistically compared with the existing transit routes
1
The optimization problem is solved using a competitive solver for MILPs (Manual, 1987) and codes are run
on a 64-bit Windows operating system with 32GB RAM and an Intel i7-3770 processor.

TRISTAN XII Symposium Original abstract submittal

 4

in Manhattan using the Kolmogorov–Smirnov test to check similarity in terms of aspects such as
length, orientation, etc. It is found that the obtained routes are not significantly different from
the current routes statistically. These results are not presented here for brevity but interested
readers can refer to Supplementary material 1 (2024). In summary, this work provides a novel
way to efficiently generate transit routes based on the demand corridors obtained from the TBA
method of Kumar et al. (2024). An important contribution of this work is that it considers
the information regarding large-scale road networks while designing transit routes in contrast to
existing studies that test their results on smaller networks.

(a) (b)

Figure 1 – The left figure shows the transit routes as obtained from the proposed formulation and
the right figure shows the existing transit routes in Manhattan, New York City.

References
Ahmed, Leena, Mumford, Christine, & Kheiri, Ahmed. 2019. Solving urban transit route design problem
using selection hyper-heuristics. European Journal of Operational Research, 274(2), 545–559.
Chakroborty, Partha, & Wivedi, Tathagat. 2002. Optimal route network design for transit systems using
genetic algorithms. Engineering optimization, 34(1), 83–100.
Dataset 1. 2014. NYC open data. https://data.cityofnewyork.us/City-Government/road/
svwp-sbcd. Accessed on 2024-09-19.
Dataset 2. 2020. MTA bus time. https://bt.mta.info/. Accessed on 2024-09-19.
Kumar, Praveen, Chakroborty, Partha, & Gehlot, Hemant. 2024. Novel Trip Agglomeration Methods for
Efficient Extraction of Urban Mobility Patterns. Networks and Spatial Economics, 1–30.
Manual, CPLEX User’s. 1987. Ibm ilog cplex optimization studio. Version, 12(1987-2018), 1.
Supplementary material 1. 2024. Distribution of parameters for transit routes. https://sites.google.
com/view/praveenkumar-iitk/supplementary-material?authuser=0. Uploaded on 2024-10-30.
Vansteenwegen, Pieter, Souffriau, Wouter, & Van Oudheusden, Dirk. 2011. The orienteering problem: A
survey. European Journal of Operational Research, 209(1), 1–10.
Vermeir, Evert, Engelen, Wouter, Philips, Johan, & Vansteenwegen, Pieter. 2021. An exact solution
approach for the bus line planning problem with integrated passenger routing. Journal of Advanced
Transportation, 2021(1), 6684795.

TRISTAN XII Symposium Original abstract submittal