Data-Driven Optimization of Public Transit

Schedule

Sanchita Basak, Fangzhou Sun, Saptarshi Sengupta and Abhishek Dubey

Department of EECS, Vanderbilt University, Nashville, TN, USA

{[email protected], [email protected],
[email protected], [email protected]}

Abstract. Bus transit systems are the backbone of public transporta-

tion in the United States. An important indicator of the quality of service
in such infrastructures is on-time performance at stops, with published
transit schedules playing an integral role governing the level of success of
the service. However there are relatively few optimization architectures
leveraging stochastic search that focus on optimizing bus timetables with
the objective of maximizing probability of bus arrivals at timepoints with
delays within desired on-time ranges. In addition to this, there is a lack of
substantial research considering monthly and seasonal variations of delay
patterns integrated with such optimization strategies. To address these,
this paper makes the following contributions to the corpus of studies on
transit on-time performance optimization: (a) an unsupervised cluster-
ing mechanism is presented which groups months with similar seasonal
delay patterns, (b) the problem is formulated as a single-objective op-
timization task and a greedy algorithm, a genetic algorithm (GA) as
well as a particle swarm optimization (PSO) algorithm are employed to
solve it, (c) a detailed discussion on empirical results comparing the al-
gorithms are provided and sensitivity analysis on hyper-parameters of
the heuristics are presented along with execution times, which will help
practitioners looking at similar problems. The analyses conducted are
insightful in the local context of improving public transit scheduling in
the Nashville metro region as well as informative from a global perspec-
tive as an elaborate case study which builds upon the growing corpus
of empirical studies using nature-inspired approaches to transit schedule
optimization.

Keywords: timetable optimization · genetic algorithm · particle swarm

optimization · sensitivity analysis · scheduling

1 Introduction
Bus systems are the backbone of public transportation in the US, carrying over
47% of all public passenger trips and 19,380 million passenger miles in the US
[18] . For the majority of cities in the US which do not have enough urban forms
or budget to build expensive transit infrastructures like subways, the reliance is
on buses as the most important transit system since bus systems have advantages
2 Basak, Sun, Sengupta and Dubey

of relatively low cost and large capacity. Nonetheless, the bus system is also one
of the most unpredictable transit modes. Our study found that the average on-
time performance across all routes of Nashville bus system was only 57.79% (see
Section 6.1). The unpredictability of delay has been selected as the top reason
why people avoid bus systems in many cities [2].
Providing reliable transit service is a critical but difficult task for all metropo-
lis in the world. To evaluate service reliability, transit agencies have developed
various indicators to quantify public transit systems through several key perfor-
mance measurements from different perspectives [4]. In the past, a number of
technological and sociological solutions have helped to evaluate and reduce bus
delay. Common indicators of public transit system evaluation include schedule
adherence, on-time performance, total trip travel time, etc. In order to track
the transit service status, transit agencies have installed AVL on buses to track
their real-time locations. However, the accuracy of AVL in urban areas is quite
limited due to the low sampling rate (every minute) and the impact of high
buildings on GPS devices. To have some basic controls during bus operation,
public transit agencies often use time point strategies, where special timing bus
stops (time points are special public transit stops where transit vehicles try to
reach at scheduled times) are deployed in the middle of bus routes to provide
better arrival and departure time synchronizations.
An effective approach for improving bus on-time performance is creating
timetables that maximize the probability of on-time arrivals by examining the
actual delay patterns. When designing schedules for real-world transport sys-
tems (e.g. buses, trains, container ships or airlines), transport planners typi-
cally adopt a tactical-planning approach [10]. Conventionally, metro transit en-
gineers analyze the historical data and adjust the scheduled time from past
experience, which is time consuming and error prone. A number of studies have
been conducted to improve bus on-time performance by reliable and automatic
timetabling. Since the timetable scheduling problem is recognized to be an NP-
hard problem [28] , many researchers have employed heuristic algorithms to solve
the problem. The most popular solutions include ad-hoc heuristic searching algo-
rithms (e.g. greedy algorithms), neighborhood search (e.g. simulated annealing
(SA) and tabu search (TS)), evolutionary search (e.g. genetic algorithm) and
hybrid search [24].
However, there are few stochastic optimization models that focus on opti-
mizing bus timetables with the objective of maximizing the probability of bus
arrivals at timepoint with delay within a desired on-time range (e.g. one minute
early and five minutes late), which is widely used as a key indicator of bus service
quality in the US [1]. A timepoint is a bus stop that is designed to accurately
record the timestamps when buses arrive and leave the stop. Bus drivers use
timepoints to synchronize with the scheduled time. For example, to quantify bus
on-time arrival performance, many regional transit agencies use the range of [-
1,+5] minutes compared to the scheduled bus stop time as the on-time standard
to evaluate bus performance using historical data [1]. The actual operation of
bus systems is vulnerable to many internal and external factors. The external
 Data-Driven Optimization of Public Transit Schedule 3

Fig. 1. The proposed toolbox for bus on-time performance optimization. City planners
use bus schedule, historical trip information and desired on-time range and layover time,
and get outputs of optimized timetable as well as estimated on-time performance.

factors include urban events (e.g., concerts, sporting events, etc.), severe weather
conditions, road construction, passenger and bicycle loading/offloading, etc. One
of the most common internal factors is the delay between two consecutive bus
trips, where the arrival delay of previous trips causes departure delay of the
next trip. Furthermore, there are monthly and seasonal variation in the actual
delay patterns, but most transit agencies publish a uniform timetable for the
next several months despite the variations. How to cluster the patterns and op-
timize timetables separately remains an open problem. Furthermore, heuristic
optimization techniques have attracted considerable attention, but finding the
optimal values of hyper-parameters are difficult, since they depend on nature
of problem and the specific implementation of the heuristic algorithms, and are
generally problem specific.

Research Contributions: In this paper, the monthly and seasonal delay

patterns are studied and outlier analysis and clustering analysis on bus travel
times to group months with similar patterns together are carried out. The fea-
ture vectors are aggregated by routes, trips, directions, timepoint segments and
months encompassing mean, median and standard deviation of the historical
travel times. This work significantly extends prior work on the problem as in
[23]. Along with a greedy algorithm and a Genetic Algorithm (GA), swarm
based optimization algorithm has been introduced in this work, where the semi-
autonomous agents in the swarm can update their status guided by the full
knowledge of the entire population state. Thus, Particle Swarm Optimization
(PSO) [14] algorithm is employed in this work to generate new timetables for
month clusters that share similar delay patterns with the goal of testing both
evolutionary computing and swarm intelligence approaches. It is observed that
the optimized on-time performance averaged across all bus routes has increased
by employing PSO as compared to that by GA. Also the execution times of
PSO are much less than GA and are more stable indicating lesser variability of
results over different runs. Sensitivity analysis on choosing the optimal hyper-
parameters for the proposed heuristic optimization algorithms are also presented.
A stability analysis of the respective algorithms have been put forward by study-
4 Basak, Sun, Sengupta and Dubey

ing the on-time performance and execution time over several runs. The overall
workflow of the proposed optimization mechanisms is illustrated in Figure 1.

The rest of the paper is organized as follows: Section 2 compares our work
with related work on transit timetabling; Section 3 presents the problem for-
mulation; Section 4 presents the details of the transit data stores; Section 5
discusses the timetable optimization mechanisms used; Section 6 evaluates the
performance of the optimization mechanisms and presents sensitivity analysis
results; Section 7 presents concluding remarks and future work.

2 Related Work and Challenges

This section compares our system with related work on transit timetable schedul-
ing. A number of studies have been conducted to provide timetabling strategies
for various objectives: (1) minimizing average waiting time [27] (2) minimizing
transfer time and cost [7][12][24], (3) minimizing total travel time [17], (4) max-
imizing number of simultaneous bus arrivals [9], [13], (5) minimizing the cost
of transit operation [26], (6) minimizing a mix of cost (both the user’s and the
operator’s) [6].
The design of timetable with maximal synchronizations of bus routes without
bus bunching has been researched by Ibarra-Rojas et al. [13]. The bus synchro-
nization strategy has been discussed from the perspective of taking waiting time
into account in the transfer stops in the work of Eranki et al. [9]. An improved
GA in minimizing passenger transfer time considering traffic demands has been
explored by Yang et al. [12]. Traffic and commuter demand has also been consid-
ered in the work by Wang et al. [27]. Other than employing optimization algo-
rithms several deep learning techniques [22] have been applied in bus scheduling
problems [15].
Nayeem et al. [17] set up the optimization problem over several criteria, such
as minimizing travel time and number of transfers and maximizing passenger
satisfaction. A route design and neighborhood search through genetic algorithm
minimizing number of transfers has been discussed by Szeto et al. [24]. Zhong
et al. [29] used improved Particle Swarm Optimization for recognizing bus rapid
transit routes optimized in order to serve maximum number of passengers.

2.1 Research Challenges

(a) Clustering Monthly and Seasonal Variations in Historical Ar-

rival Data: Studying the historical travel time at segments can be an effective
way to set bus timetables. However, existing work doesn’t consider the monthly
and seasonal variation in historical monthly data, and the variation can be uti-
lized for better scheduling. Generating one timetable for all months may not be
the best solution. As traffic and delay patterns are prone to changes over seasonal
variations and various times, we generate clusters grouping months with unsuper-
vised algorithm and develop optimization strategies for the generated clusters.
 Data-Driven Optimization of Public Transit Schedule 5

Table 1. The scheduled time and recorded actual arrival and departure time of two
sequential trips that use the same bus of route 4 on Aug. 8, 2016. The arrival delay at
the last timepoint of the first trip accumulates at the first timepoint of the second trip.

Timepoints
MCC4 14 SY19 PRGD GRFSTATO
Scheduled Time 10:50 AM 11:02 AM 11:09 AM 11:18 AM
Trip 1 Actual Arrival Time 10:36 AM 11:10 AM 11:18 AM 11:27 AM
Actual Departure Time 10:50 AM 11:10 AM 11:18 AM 11:30 AM
Scheduled Time 11:57 AM 11:40 AM 11:25 AM 11:20 AM
Trip 2 Actual Arrival Time 12:11 PM 11:51 AM 11:34 AM 11:27 AM
Actual Departure Time 12:11 PM 11:51 AM 11:34 AM 11:30 AM

We evaluate the proposed mechanism via simulation. The cluster-specific sched-

ule is shown to further increase the on-time performance compared to generating
one uniform timetable.

(b) Computing Efficiently and Accurately in the Solution Space:

Transit performance optimization techniques rely on historical delay data to set
up new timetables. However, the large amount of historical data makes it a chal-
lenge to compute efficiently. For example, Nashville MTA updates the bus sched-
ule every 6 months but each time there are about 160,000 historical record entries
to use. Moreover, the solution space has typically very large under constraints
(e.g., sufficient dwelling time at bus stops, adequate layover time between trips,
etc.). A suitable optimization algorithm is necessary for efficient and accurate
computation. Since this is a discrete-variable optimization problem, gradient-
based methods cannot be used and gradient-free methods need to be considered.
A naive algorithm for discrete optimization is exhaustive search, i.e., every fea-
sible time is evaluated and the optimum is chosen. Exhaustive search works for
a small finite number of choices, and cannot be used for high-dimensional prob-
lems. Genetic algorithm [7][6], as well as particle swarm optimization [14] are
used commonly in solving heuristic problems . Thus we consider applying ge-
netic algorithm and particle swarm optimization (PSO) in the context. Section 5
describes the key steps of how we apply greedy, genetic and PSO algorithms to
solve the timetable optimization problem.

3 Problem Formulation

Typically, transit delay are not only affected by external factors (such as traffic,
weather, travel demand, etc.), but also by some internal factors. For example,
the accumulated delay occurred on previous trips may cause a delay in consec-
utive trips by affecting the initial departure time of the next trip. In order to
illustrate the problem context with simplicity and without generality, we take
two sequential bus trips of route 4 in Nashville as an example (the scheduled
time and the actual arrival and departure time recorded on Aug. 8, 2016 are
6 Basak, Sun, Sengupta and Dubey

Fig. 2. (a) A route segment on bus route 3 leaving downtown; (b) The variance of
actual travel time and (c) the relative standard deviation of actual travel times on a
bus route segment in time period between Sept. 1, 2016 and Feb. 28, 2017.

shown in Table 1) to describe the optimization problem. On each service day,

after a vehicle of the first trip (121359) arrives at the last stop (Timepoint GRF-
STATO) with scheduled time of 11:18 AM, the second trip is scheduled to depart
using the same vehicle from the same stop at 11:20 AM. On Aug. 8, 2016, the
arrival time at the last stop (Timepoint GRFSTATO) of the first trip (121359)
is exceptionally late for 9 minutes, which contributes to the 10-minute depar-
ture delay at the beginning of the second trip. Since the scheduled layover time
between the two trips is only 2 minutes (between 11:18 AM and 11:20 AM), any
large delay at the first trip is very likely to transfer to the next trip. Therefore,
the optimization problem should involve a process that considers not only the
travel delay on segments, but also the improper lay over time between trips.
Figure 2 illustrates the large variation of bus travel time distribution. The
example shows travel time data collected from bus trips depart at a specific time
of the day on route 3 in Nashville. The coefficient of variation (also known as
relative standard deviation) , which is a standardized measure of dispersion of
a probability distribution, is very high on all timepoints along the route. The
complexity and uncertainty of travel times introduce great challenges to the task
of timetable optimization.

3.1 Problem Definition

For a given bus trip schedule b, let H = {h1 , h2 , ..., hm } be a set of m his-
torical trips with each trip passing n timepoints {s1 , s2 , ..., sn }. So the on-time
performance of the bus trip schedule b can be defined as a ratio of an indica-
tor function I(hi , sj ) summed over all timepoints for all historical trips to the
product of the total number of historical trips and total number of timepoints.
The indicator function I(hi , sj ) is 1 if di,j ∈ [tearly , tlate ], otherwise 0, where
di,j = tarrival
hi ,sj − Tharrival
i ,sj
 Data-Driven Optimization of Public Transit Schedule 7

The objective is to design a schedule optimization problem to generate new

departure
Th,s , ensuring on-time performance maximization. tearly and tlate are two
time parameters pre-defined by the transit authority as a measure of schedule
maintenance and di,j is the actual delay that arriving in timepoint sj .

4 Data Store
4.1 Data Sources
We established a cloud data store and reliable transmission mechanisms to feed
our Nashville Metropolitan Transit Authority (MTA) updates the bus schedule
information every six months and provides the schedule to the public via GTFS
files. In order to coordinate and track the actual bus operations along routes,
MTA has deployed sensor devices at specially bus stops (called timepoints) to
accurately record the arrival and departure times. In Nashville, there are over
2,700 bus stops all over the city and 573 of them are timepoint stops. City
planners and MTA engineers analyze the arrival and departure records regularly
to update the transit schedule. The details of the datasets are as follows:

– Static GTFS. This dataset defines the static information of bus schedule
and associated geographic information, such as routes, trips, stops, depar-
ture times, service days, etc. The dataset is provided in a standard transit
schedule format called General Transit Feed Specification (GTFS).
– GTFS-realtime. This dataset is recorded real-time transit information in
GTFS-realtime format, which include bus locations, trip updates and service
alerts. The GTFS-realtime feed is collected and stored in one-minute interval.
– Timepoints. This dataset provides accurate and detailed historical arrival
and departure records at timepoint stops. The information include route,
trip, timepoint, direction, vehicle ID, operator, actual arrival and departure
time, etc. The dataset is not available in real-time but collected manually
by Nashville MTA at the end of each month.

Even though the same timepoint datasets are utilized in the study, the pro-
posed method is not limited to the timepoint datasets and can use some surro-
gate data sources: (1) automatic passenger counters (APC) data: APC datasets
records both passenger counts and departure/arrival times at stops (2) GTFS-
realtime feed: the real-time bus locations reported by automatic vehicle locator
(AVL) installed on buses. Compared with timepoint datasets, APC data also
provides accurate times at normal stops thus it is the most suitable alternative
dataset. However, GTFS-realtime suffers from low sampling rate and low accu-
racy in the city and may reduce the performance of the proposed mechanism.

4.2 Data Cleaning

Since raw transit dataset often contains missing, duplicate and erroneous sam-
ples, preprocessing is a necessary step to prepare a clean and high-quality dataset.
8 Basak, Sun, Sengupta and Dubey

Missing data issue occurs due to hardware or network problems. Generally,

there are samples with missing data can be dropped or filled with a specific or
average values. Duplicated data (e.g., a bus trip is recorded more than one time)
will oversample certain delay values and make the delay dataset biased. We drop
the trips with no historical records and remove duplicated records.
Outliers are values that are distant from most of the observed data in presence
of which clustering can be inappropriate. K-means clustering algorithm is also
sensitive to outliers present in the data. The approach taken here is to calculate
Median Absolute Deviation (MAD), a robust measure of statistical dispersion.
The MAD of a data set [X = (x1 , x2 , ..., xn )] can be calculated as: M AD =
median(|xi −median(X)|). For normal distribution the scaled MAD is defined as
(MAD/0.6745), approximately equal to the standard deviation. xi is considered
an outlier if the difference between xi and median is larger than three times of
standard deviation (i.e. scaled MAD).

5 Timetable Optimization Mechanisms

5.1 Month Grouping by Clustering Analysis
This section introduces a clustering analysis mechanism that groups months
with similar transit delay patterns together and the results will later be used to
generate separate timetables for each group.
Feature Engineering. We assume the monthly delay patterns can be repre-
sented by the mean, median and standard deviation that derived from historical
delay data. Considering a bus trip consists of n timepoints, there are n − 1 seg-
ments between the timepoints. The mean value µ, the median value m, and the
standard deviation σ of the historical travel times for each timepoint segment in
each month are integrated to generate feature vectors to represent the historical
delay data distribution:

[µ1 , m1 , σ1 , µ2 , m2 , σ2 , ..., µn−1 , mn−1 , σn−1 ] (1)

Month Clustering. Clustering is an unsupervised/supervised learning tech-

nique for grouping similar data. We employ k-means algorithms to identify the
homogeneous groups where months share similar patterns. The trip data per
month is first normalized and then clustered using feature vectors (in Equa-
tion 1) by K-Means algorithm:
k X
X
arg min kx − µi k2 (2)
S i=1 x∈Si

where µi is the mean of all datapoints in cluster Si . Determining the optimal

number of clusters in a data set is a fundamental issue in partitioning clustering.
For k-means algorithms, the number of clusters is a hyper-parameter that needs
to be set manually. An upper bound is set in advance. Elbow [16], Silhouette [19]
and gap statistic [25] methods are popular direct and statistical methods to find
 Data-Driven Optimization of Public Transit Schedule 9

Fig. 3. The feature vectors [mean, standard deviation, median] of the travel time in 4
months of 2016 for a segment (WE23-MCC5 5) on a bus trip of route 5.

the optimal number of clusters. Particularly, Silhouette analysis is employed in

this study to measure how close each point is to others within one cluster. The
silhouette score s(i) is defined as:
b(i) − a(i)
s(i) = (3)
max{a(i), b(i)}
where for each data point with index i in the cluster, ai is the average distance
between datai and the rest of data points in the same cluster, bi is the smallest
average distance between datai and every other cluster. Some other clustering
techniques that can be applied to these kind of problem can be found in [20].
Example. Figure 3 plots the [mean, standard deviation, median] vectors of
the monthly travel time for a segment (WE23-MCC5 5) on a bus trip of route
5 (Figure 6(a)). From figure 3 the monthly variation of the data is evident and
hence two clusters ([May, June, July] and [August]) can be formed from these
four months of data to prepare distinct schedules for the clusters.

5.2 Estimating On-time Performance of Transit Schedules

Historical Dwell Time Estimation. Travel demand at bus stops is important
statistics for setting up proper schedule times. However, for bus systems without
automatic passenger counters (APCs), historical travel demand (represented by
number of commuters boarding) is not available in original datasets. To get
demand patterns, we utilize historical arrival and departure times to estimate
the dwell time caused by passengers. Particularly, we consider the following
two scenarios in historical records: (1) when a bus turns up at a stop, earlier
than scheduled time, the waiting time between the scheduled time and actual
departure time is used, (2) on the other hand, when it turns up later than
scheduled time, the waiting time between the actual arrival time and departure
time is used. As shown in Table 1, for the timepoint SY19 on trip 1 with scheduled
time of 11:02 AM:
– For the case when a bus arrived earlier at 10:58 AM instead of the scheduled
time at 11:02 AM and departed at 11:04 AM, as the bus would always wait
10 Basak, Sun, Sengupta and Dubey

there at least for 4 minutes (the difference between the actual and scheduled
arrival time) irrespective of presence of passengers, the dwell time caused by
passengers is calculated as the additional time taken for departure after the
scheduled time (11:04 AM - 11:02 AM = 2 minutes).
– On the other hand, if the bus arrived later at 11:05 AM and departed at
11:06 AM, then the dwell time caused by passengers is calculated as the
additional time spent after the actual arrival time (11:06 AM - 11:05 AM =
1 minutes).
Arrival Time Estimation. The arrival time of a bus at a stop is impacted by
two factors: (1) travel times at segments before the stop, and (2) dwell times at
the previous stops. We assume that a bus will wait until the scheduled time if it
arrives earlier than the scheduled time, and the historical travel time between two
timepoints will remain the same in the simulation. In order to obtain an estimate
of the arrival time, the historical dwell time caused by commuters (which in
turn is representative of the historical travel demand), is factored into account
by adding it to the arrival time at any timepoint. The simulation will stall for
an additional time till the new scheduled time is reached in the event that the
previous sum is earlier than the new scheduled time. By taking into consideration
the simulated departure time stdepart
h,sj at previous timepoint sj , the actual travel
time tarrive
sj+1 − tdepart
sj between sj and sj+1 , the dwell time tdwell
sj+1 , the simulated
departure time stdepart
h,sj+1 at a timepoint sj+1 can be found out. The new schedule
depart
time Th,ss j+1
at sj+1 is expressed as:

stdepart depart depart

h,sj+1 = max(Th,ssj+1 , sth,sj + (tarrive depart
sj+1 − tsj ) + tdwell
sj+1 ) (4)

5.3 Timetable Optimization Using a Greedy Algorithm

We employed a greedy algorithm that adjusts the scheduled arrival time greed-
ily and sequentially for the succeeding segments between timepoints. The main
objective is to optimize the bus arrival time for succeeding timepoints such that
new optimized schedule is guaranteed to maximize the probability of bus arrivals
between any two consecutive stops with delay bounded in the desired range of
[tearly , tlate ].
We utilized the empirical cumulative distribution function (CDF) to evaluate
the percentage of historical delay in desired range instead of assuming that the
data is drawn from any specific distribution (e.g. Gaussian distribution).
An empirical CDF is a non-parametric estimator of the CDF of a random
variable. The empirical CDF of variable x is defined as:
n
X
F̂n (x) = P̂n (X ≤ x) = n−1 I(xi ≤ x) (5)
n=1

where I() is an indicator function:

(
1, if xi ≤ x
I(xi ≤ x) = (6)
0, otherwise
 Data-Driven Optimization of Public Transit Schedule 11

Then the CDF of x in range [x + tearly , x + tlate ] can be calculated using the
following equation:

F̂n (x + tlate ) − F̂n (x + tearly )

n
X (7)
= n−1 I(x + tearly ≤ xi ≤ x + tlate )
n=1

5.4 Timetable Optimization Using Heuristic Algorithms

The performance optimization for scheduling transit vehicles is a multidimen-

sional problem and as such the objective function is nonconvex in nature con-
sisting of several troughs and ridges. Hence, to compute the optimally scheduled
routing strategy with acceptable time constraints, an approach powered by high
quality of solution estimation techniques such as evolutionary algorithms and
metaheuristics can be considered.

Genetic Algorithm Genetic algorithm [11] is a heuristic optimization al-

gorithm that derives from biology. The basic steps involved in genetic algo-
rithms include initialization, selection, crossover, mutation, and termination. The
timetable for each trip is decided by the scheduled departure time at the first
stop as well as the scheduled travel time between any two subsequent timepoints
along the trip. Since our goal is to update timetables to make the bus arrivals
more on time, we assign the scheduled travel times between timepoints as chro-
mosomes in populations, and use the on-time performance estimation mechanism
proposed in Section 5.2 as objective functions. The chromosome of the individual
solutions in the genetic algorithm is a vector of integers representing travel time
between subsequent timepoints. In order to reduce the search space and match
the real-world scenarios, the travel time in each individual is re-sampled to a
multiple of 60 seconds and restricted to the unit of minutes. The performance of
this algorithm is governed by different hyperparameters such as population size,
crossover and mutation rate controlling the algorithm’s exploitation and explo-
ration capability. The choice of such hyperparameters are explained in detail in
section 6.

Particle Swarm Optimization Eberhert and Kennedy [14] proposed parti-

cle swarm optimization (PSO) as a stochastic population based optimization
algorithm which can work with non-differentiable objective function without ex-
plicitly assuming its underlying gradient disparate from gradient descent tech-
niques. The interested reader is directed to [21] by Sengupta et al. for a detailed
understanding of the algorithm. PSO has been shown to satisfactorily provide
solutions to a wide array of complex real-life engineering problems, usually out
of scope of deterministic algorithms [8][5][3]. PSO exploits the collective intelli-
gence arising out of grouping behavior of flocks of birds or schools of fish.This
manifestation of grouping is termed as ’emergence’, a phenomenon in which a
12 Basak, Sun, Sengupta and Dubey

cohort of individuals from a social network is aimed to accomplish a task be-

yond their individual capability. Likewise, each particle in the swarm, represents
a potential solution to the multi-dimensional problem to be optimized.
Initialization Each particle has certain position which can be thought of
as a collection of co-ordinates representing the particle’s existence in a specific
region in the multidimensional hyperspace. As a particle is a potential solution
to the problem, the particle’s position vector has the same dimensionality as the
problem. The velocity associated with each particle is the measure of the step
size and the direction it should move in the next iteration.
Each particle in the swarm maintains an n-dimensional vector of travel times.
At first, the position for each particle in the population is initialized with the set
of travel time between the timepoints randomly selected between the minimum
and maximum of the aggregated actual historical data. With swarm size as p, ev-
ery particle i (1<i <p) maintains a position vector xi =(xi1 ,xi2 ,xi3 ,...,xin ) and a
velocity vector vi =(vi1 ,vi2 ,vi3 ,...,vin ) and a set of personal bests pi =(pi1 ,pi2 ,pi3 ,...,pin ).
Optimization At each iteration, the position of a particle is updated, and
compared with the personal best (pbest) obtained so far. If the fitness due to
the position obained at current iteration is more (as it is a fitness maximization
problem) than the pbest obtained upto the previous iteration, then the current
position becomes the personal best or pbest, otherwise pbest remains unchanged.
Thus the best position of a particle obtained so far is stored as pbest. The global
best or gbest is updated when the population’s overall current best, i.e., the best
of the pbsests is better than that found in the previous iteration.
After initializing positions and velocities, each particle updates its velocity
based on previous velocity component weighted by an inertial factor , along with
a component proportional to the difference between its current position and pbest
weighted by a cognition acceleration coefficient, and another component propor-
tional to the difference between its current position and (gbest), weighted by a
social acceleration coefficient. This is socio-cognitive model of PSO and facili-
tates information exchange between members of the swarm. Since all members
are free to interact with each other, the flow of information is unrestricted and
the PSO algorithm is said to have a ’fully-connected’ topology. While updating
the velocity, a particle’s reliance on its own personal best is dictated by its cog-
nitive ability, and the reliance on the entire swarm’s best solution is dictated
by its social interactive nature. Hence those factors in the velocity component
are weighted by the cognition acceleration coefficient c1 and social acceleration
coefficient c2. The new positions of the particles are updated as the vector sum
of the previous positions and the current velocities. Thus the positions of the
particles, are updated aiming towards intelligent exploration of the search space,
and subsequent exploitation of the promising regions in order to find the optimal
solution based on fitness optimization of the stated problem.
After each iteration is completed, the velocity and position of a particle are
updated as follows:
vi,j (t + 1) = w.vi,j (t) + c1 .r1 (t).(pi,j (t) − xi,j (t)) + c2 .r2 (t).(pg,j (t) − xi,j (t)) (8)
xi,j (t + 1) = xi,j (t) + vi,j (t + 1) (9)
 Data-Driven Optimization of Public Transit Schedule 13

ALGORITHM 1: Particle Swarm Optimization algorithm for bus on-time perfor-

mance optimization
Data: D ← Historical timepoint datasets
Input : (1) [tearly ,tlate ] ← on-time range , (2) maxIter ← maximum number of
iterations maxIter, (3) npop ← number of particles in the population size
npop, (4) w ← inertia weight, (5) c1 ← cognition acceleration coefficient,
(6) c2 ← social acceleration coefficient, (7) h ← bus trip for optimization,
(8) upperLimit ← upper limit of the number of clusters
Output: Optimized schedule b at timepoints for bus trip h
GetAllTimepoints(D, h);
GetHistoricalData(D, h);
monthClusters ← ClusterMonthData(upperLimit);
for monthCluster ∈ monthClusters do
P ← [];
for population size npop do
Initialize each particle with random position and velocity
P ← P ∪ InitialIndividual();
end
while maxIter is not reached do
Evaluate the fitness function (J ) for each particle’s position (x )
if J(x) ¿ J(pbest), then pbest = x
gbest ← Update if the population’s overall current best is better than that in
previous iteration
Update the velocity of each particle according to equation (10)
Update the position of each particle according to equation (11)
end
Give gbest as the optimal schedule b at timepoints for bus trip h
end

vi,j and xi,j represent the velocity and position of the i-th particle in the j-th
dimension. Cognition and social acceleration coefficients are indicated by c1 and
c2 , whereas r1 and r2 are random numbers uniformly distributed between 0 to
1. pi,j represents a particle0 s personal best and pg,j represents the global best
of the population. w acts as an inertial weight factor controlling the exploration
and exploitation of new positions in the search space and t denotes the number
of iterations.
The problem is formulated as fitness maximization problem in order to bring
out optimal travel times to improve on-time performance. Hence the personal
best of a particle is updated as follows at the end of each iteration.

(
pi,j (t), if f itness(xi,j (t + 1)) ≤ f itness(pi,j (t))
pi,j (t + 1) = (10)
xi,j (t + 1), if f itness(xi,j (t + 1)) > f itness(pi,j (t))

Termination The termination condition set for PSO is the predefined max-
imum number of iterations. Since the optimized on time performance is different
for each trip, the termination condition is not set as any predefined upper limit
14 Basak, Sun, Sengupta and Dubey

Table 2. Comparison of original and optimized on-time performance averaged across

all bus routes for GA without and with clustering and PSO with clustering respectively.

Original GA w/o. Clustering GA w. Clustering PSO w. Clustering

On-time Perf. 57.79% 66.24% 68.34% 68.93%

of the fitness value. With other hyperparameters fixed PSO can produce the
optimal solution approximately in 30 iterations for this problem.
The pseudo code for PSO is discussed in Algorithm 3. Historical timepoint
datasets are used to conduct the particle swarm optimization algorithm for this
problem. The input includes on-time range, maximum number of iterations,
number of particles in the population size, inertia factor, cognition and social
acceleration coefficient, bus trip and upper limit of number of month clusters.

6 Evaluation of the Results

6.1 Evaluating the Clustering Analysis
To evaluate the effectiveness of the clustering analysis, we compared the op-
timized on-time performance with and without a clustering analysis step: (1)
months are not clustered and a single timetable is generated for all months, (2)
month clustering is conducted at first and the optimization algorithms is applied
on different month groups to generate separate timetables.
Table 2 shows the original and optimized on-time performance on average
across all bus routes. Using the genetic algorithm without clustering step im-
proved the original performance from 57.79% to 66.24%. By adding the clustering
step which groups months with similar patterns the performance was improved
to 68.34%.

Fig. 4. The chart shows the simulation results of on-time performance and execution
times for GA and PSO to run 10 times.
 Data-Driven Optimization of Public Transit Schedule 15

6.2 Comparing Optimization Performance of Greedy, Genetic and

PSO Algorithms

The original on-time performance, optimized on-time performance using greedy

algorithm, genetic algorithm and PSO are illustrated in Figure 5. It is observed
that while all the algorithms can improve the on-time performance, the genetic
algorithm and PSO outperforms the greedy algorithm because they optimize
the schedule for all timepoint segments on each trip all together. The original
on-time performance of all bus routes is 57.79%. The greedy algorithm improved
it to 61.42% and the genetic algorithm improved it further to 68.34%. The PSO
algorithm has a slightly better optimized on-time performance of 68.93%. Fig-
ure 4 shows the simulation results of the stability analysis for GA and PSO. Even
though GA and PSO got similar on-time performance, with PSO surpassing the
performance of GA by a small extent, the execution times of PSO are much less
than GA and are more stable.

Fig. 5. The original on-time performance and the optimized on-time performance using
greedy algorithm, genetic algorithm with/without clustering analysis and PSO algo-
rithm.

6.3 Sensitivity Analysis on the Hyper-parameters of the Genetic

algorithm

We designed three simulations that choose different hyper-parameters: (1) pop-

ulation sizes that range from 10 to 110, (2) crossover rates that range from 0.1
to 1.0, (3) mutation rates that range from 0.1 to 1.0. Real-world data is col-
lected from Route 5, which is one of major bus routes that connects downtown
Nashville and the southwest communities in Nashville. The route contains 6
16 Basak, Sun, Sengupta and Dubey

timepoint stops and 5 segments between the 6 timepoint stops. The bus trips
with direction from Downtown are selected. The goal is to maximize the on-time
performance for these trips by optimizing the schedule time at the 6 timepoint
stops.
Figure 6(b) shows the simulation results of choosing different population
sizes. Increasing the population size from 10 to 90 results a better on-time perfor-
mance, however, increasing the size ever further doesn’t help making the on-time
performance any better. On the other hand, the total time increases linearly as
the population size grows. So a population size around 90 is the optimal size to
use.
Figure 6(c) illustrates results of using different crossover rates. The optimized
on-time performance remains almost the same for the crossover range, but there
is a significant difference in terms of the total execution time. The crossover rate
impacts the exploitation ability. A proper crossover rate in the middle of the
range can faster the process to concentrate on an optimal point.
Figure 6(d) show the simulation results when using different mutation rates.
The total execution time is small when the mutation rate is either very small or
very large. Mutation rates controls the exploration ability. During the optimiza-
tion, a small mutation rate will make sure the best individuals in a population do
not vary too much in the next iteration and thus is faster to get stable around the
optimal points. So we suggest setting a very small mutation rates when running
the proposed algorithm.

6.4 Sensitivity Analysis on the Hyper-parameters of Particle

Swarm Optimization

We designed four simulation setups that choose different hyper-parameters: (1)

The inertial weight factor,w that range from 1 to 8, (2) Social acceleration coef-
ficient c1 that range from 1 to 8, (3) Cognition acceleration coefficient c2 that
range from 1 to 8, and (4) Number of particles that range from 2 to 36. Real-
world data regarding bus timings is collected from Route 8, which is one of the
major bus routes that connects Music City Central Nashville and the Lipscomb
University in Nashville. The route contains 5 timepoint stops and 4 segments
between the 5 timepoint stops. The goal is to maximize the on-time performance
for these trips by optimizing the schedule time at the 5 timepoint stops.
Figure 6(e) shows the simulation result while optimizing for the inertial
weight w by varying it. It is observed that the optimized on-time performance
is at its peak when w is nearly equal to 5 with less execution time. Performance
deteriorates along with an increase in execution time as the selection is moved
away from 5. So, an optimal value to choose for w, will be somewhere around 5.
Figure 6(f) shows the simulation result for optimizing the cognition accelera-
tion coefficient c1 by varying it. The particle has a velocity component towards
its own best position weighted by c1, hence the term ’cognitive’. It is observed
that the optimized on-time performance is improved when c1 increases from 3
to 5 with less execution time. After that the performance deteriorates along with
 Data-Driven Optimization of Public Transit Schedule 17

Fig. 6. (a) Timepoints on bus route 5 in Nashville [23], (b) The the on-time perfor-
mance and overall execution time for different population sizes for GA, (c) The on-time
performance and overall execution time for different crossover rates, which controls the
exploitation ability of the GA, (d) The on-time performance and overall execution time
for different mutation rates, which controls the exploration ability of the GA, (e) The
on-time performance and overall execution time for different inertia weights, exploring
new regions of search space in PSO, (f) The on-time performance and overall execution
time for different cognition acceleration coefficients c1, in PSO, (g) The on-time per-
formance and overall execution time for different social acceleration coefficients c2, in
PSO, (h) The on-time performance and overall execution time for various population
size, in PSO
18 Basak, Sun, Sengupta and Dubey

increase in execution time. So, an optimal value to choose for c1, will be within
the range specified.
Figure 6(g) shows the simulation result for optimizing for the social accelera-
tion coefficient c2 by varying it. The particle has a velocity component towards
the global best position weighted by c2, hence the term social. It is observed that
the optimized on-time performance is improved when c2 is equal to 5 with less
execution time. Also, c2 being 4 produces good results, but there is an increase
in execution time at that value. But the overall effect of parameter c2 affects
the on-time performance only within a range of two percent. Sometimes, PSO
is able to produce optimal or near optimal performance, when all other hyper-
parameters are fixed, and thus is not sensitive to a particular hyperparameter,
which is the case considered here. So an optimal value to choose for c2, may be
close to 5, maintaining approximately a ratio near to 1:1:1 among w, c1 and c2.
Figure 6(h) shows the simulation result for optimizing the number of parti-
cles by varying the population size. It is observed that the optimized on-time
performance is maximized when the number of particles reaches 30. The execu-
tion time increases with the number of particles, so it is better to choose such
number of particles that produces the best pair in the accuracy-execution time
tradeoff. So, the population size can be chosen as 30 in this case as it yields
equally efficient results with a relatively small execution time.
Although a good insight about choice of hyperparameters can be obtained
from this sensitivity analysis, variations of the hyperparameters may produce
better results in specific routes.

7 Conclusion

In this paper, we presented research findings within a bus on-time performance

optimization framework that significantly extends our prior work [23] by propos-
ing a stochastic optimization toolchain and presenting sensitivity analyses on
choosing optimal hyper-parameters. Particularly, we describe an unsupervised
analysis mechanism to find out how months with similar delay patterns can be
clustered to generate new timetables. A classical, fully-connected PSO is bench-
marked against a greedy algorithm as well as a genetic algorithm in order to
optimize the schedule time to maximize the probability of bus trips that reach
the desired on-time range. It is observed that the PSO implementation improves
the bus on-time performance compared to other heuristics while requiring lesser
execution time across all routes. Simulations of optimization performance as well
as sensitivity analyses on the hyper-parameters of the GA and PSO algorithms
are conducted. The results indicate different strategies for choosing between the
genetic algorithm and PSO, and selecting optimal hyper-parameters guided by
the problem specificity and resource availability. With the knowledge of this
extensive study on applying guided random search techniques for bus on-time
performance optimization and the selection of hyperparameters that generate
promising results, a possible extension of this generalizable architecture to other
real-world optimization problems is worth looking at as future work.
 Data-Driven Optimization of Public Transit Schedule 19

Acknowledgments
This work is supported by The National Science Foundation under the award
numbers CNS-1528799 and CNS-1647015 and 1818901 and a TIPS grant from
Vanderbilt University. We acknowledge the support provided by our partners
from Nashville Metropolitan Transport Authority.

References
1. Arhin, S.A., Noel, E.C., Dairo, O.: Bus stop on-time arrival performance and cri-
teria in a dense urban area. International Journal of Traffic and Transportation
Engineering 3(6), 233–238 (2014)
2. Association, A.P.T.: Ridership report archives (2017)
3. Banks, A., Vincent, J., Anyakoha, C.: A review of particle swarm optimiza-
tion. part ii: hybridisation, combinatorial, multicriteria and constrained optimiza-
tion, and indicative applications. Natural Computing 7(1), 109–124 (Mar 2008).
https://doi.org/10.1007/s11047-007-9050-z, https://doi.org/10.1007/s11047-007-
9050-z
4. Benn, H.: Bus route evaluation standards, transit cooperative research program,
synthesis of transit practice 10. Transportation Research Board, Washington, DC
(1995)
5. Bouyer, A., Hatamlou, A.: An efficient hybrid clustering method
based on improved cuckoo optimization and modified particle
swarm optimization algorithms. Applied Soft Computing 67, 172 –
182 (2018). https://doi.org/https://doi.org/10.1016/j.asoc.2018.03.011,
http://www.sciencedirect.com/science/article/pii/S1568494618301273
6. Chakroborty, P.: Genetic algorithms for optimal urban transit network design.
Computer-Aided Civil and Infrastructure Engineering 18(3), 184–200 (2003)
7. Chakroborty, P., Deb, K., Subrahmanyam, P.: Optimal scheduling of urban transit
systems using genetic algorithms. Journal of transportation Engineering 121(6),
544–553 (1995)
8. Dhabal, S., Sengupta, S.: Efficient design of high pass fir filter using quantum-
behaved particle swarm optimization with weighted mean best position. In:
Proceedings of the 2015 Third International Conference on Computer, Com-
munication, Control and Information Technology (C3IT). pp. 1–6 (Feb 2015).
https://doi.org/10.1109/C3IT.2015.7060145
9. Eranki, A.: A model to create bus timetables to attain maximum synchronization
considering waiting times at transfer stops (2004)
10. Fan, W., Machemehl, R.B.: Optimal transit route network design problem with
variable transit demand: genetic algorithm approach. Journal of transportation
engineering 132(1), 40–51 (2006)
11. Goldberg, D.E.: Genetic Algorithms in Search, Optimization and Machine Learn-
ing. Addison-Wesley Longman Publishing Co., Inc., Boston, MA, USA, 1st edn.
(1989)
12. Hairong, Y., Dayong, L.: Optimal regional bus timetables using improved genetic
algorithm. In: Intelligent Computation Technology and Automation, 2009. ICI-
CTA’09. Second International Conference on. vol. 3, pp. 213–216. IEEE (2009)
13. Ibarra-Rojas, O.J., Rios-Solis, Y.A.: Synchronization of bus timetabling. Trans-
portation Research Part B: Methodological 46(5), 599–614 (2012)
20 Basak, Sun, Sengupta and Dubey

14. Kennedy, J., Eberhart, R.: Particle swarm optimization. In: Neural Networks, 1995.
Proceedings., IEEE International Conference on. vol. 4, pp. 1942–1948 vol.4 (Nov
1995). https://doi.org/10.1109/ICNN.1995.488968
15. Khiari, J., Moreira-Matias, L., Cerqueira, V., Cats, O.: Automated setting of bus
schedule coverage using unsupervised machine learning. In: Bailey, J., Khan, L.,
Washio, T., Dobbie, G., Huang, J.Z., Wang, R. (eds.) Advances in Knowledge
Discovery and Data Mining. pp. 552–564. Springer International Publishing, Cham
(2016)
16. Kodinariya, T.M., Makwana, P.R.: Review on determining number of cluster in
k-means clustering. International Journal 1(6), 90–95 (2013)
17. Nayeem, M.A., Rahman, M.K., Rahman, M.S.: Transit network design by genetic
algorithm with elitism. Transportation Research Part C: Emerging Technologies
46, 30–45 (2014)
18. Neff, J., Dickens, M.: 2016 public transportation fact book (2017)
19. Rousseeuw, P.J.: Silhouettes: a graphical aid to the interpretation and validation
of cluster analysis. Journal of computational and applied mathematics 20, 53–65
(1987)
20. Sengupta, S., Basak, S., Peters, R.A.: Data clustering using a hybrid of fuzzy
c-means and quantum-behaved particle swarm optimization. In: 2018 IEEE 8th
Annual Computing and Communication Workshop and Conference (CCWC). pp.
137–142 (Jan 2018). https://doi.org/10.1109/CCWC.2018.8301693
21. Sengupta, S., Basak, S., Peters, R.A.: Particle swarm optimization: A survey of his-
torical and recent developments with hybridization perspectives. Machine Learn-
ing and Knowledge Extraction 1(1), 157–191 (2018), http://www.mdpi.com/2504-
4990/1/1/10
22. Sengupta, S., Basak, S., Saikia, P., Paul, S., Tsalavoutis, V., Atiah, F., Ravi,
V., Peters, R.A.: A review of deep learning with special emphasis on ar-
chitectures, applications and recent trends. CoRR abs/1905.13294 (2019),
http://arxiv.org/abs/1905.13294
23. Sun, F., Samal, C., White, J., Dubey, A.: Unsupervised mechanisms for optimiz-
ing on-time performance of fixed schedule transit vehicles. In: Smart Computing
(SMARTCOMP), 2017 IEEE International Conference on. pp. 1–8. IEEE (2017)
24. Szeto, W.Y., Wu, Y.: A simultaneous bus route design and frequency setting prob-
lem for tin shui wai, hong kong. European Journal of Operational Research 209(2),
141–155 (2011)
25. Tibshirani, R., Walther, G., Hastie, T.: Estimating the number of clusters in a
data set via the gap statistic. Journal of the Royal Statistical Society: Series B
(Statistical Methodology) 63(2), 411–423 (2001)
26. Ting, C.J., Schonfeld, P.: Schedule coordination in a multiple hub transit network.
Journal of urban planning and development 131(2), 112–124 (2005)
27. Wang, Y., Zhang, D., Hu, L., Yang, Y., Lee, L.H.: A data-driven and optimal bus
scheduling model with time-dependent traffic and demand. IEEE Transactions on
Intelligent Transportation Systems 18(9), 2443–2452 (2017)
28. Wu, Y., Yang, H., Tang, J., Yu, Y.: Multi-objective re-synchronizing of bus
timetable: Model, complexity and solution. Transportation Research Part C:
Emerging Technologies 67, 149–168 (2016)
29. Zhong, S., Zhou, L., Ma, S., Jia, N., Zhang, L., Yao, B.:
The optimization of bus rapid transit route based on an im-
proved particle swarm optimization. Transportation Letters 10(5),
257–268 (2018). https://doi.org/10.1080/19427867.2016.1258972,
https://doi.org/10.1080/19427867.2016.1258972