Turkish Journal of Electrical Engineering and Computer

Sciences

Volume 32 Number 1 Article 1

2-7-2024

Traffic signal optimization using multiobjective linear

programming for oversaturated traffic conditions
Mustafa Murat COŞKUN

Cevat ŞENER

İsmail Hakkı TOROSLU

Follow this and additional works at: https://journals.tubitak.gov.tr/elektrik

Part of the Computer Engineering Commons, Computer Sciences Commons, and the Electrical and
Computer Engineering Commons

Recommended Citation
COŞKUN, Mustafa Murat; ŞENER, Cevat; and TOROSLU, İsmail Hakkı (2024) "Traffic signal optimization
using multiobjective linear programming for oversaturated traffic conditions," Turkish Journal of Electrical
Engineering and Computer Sciences: Vol. 32: No. 1, Article 1. https://doi.org/10.55730/1300-0632.4052
Available at: https://journals.tubitak.gov.tr/elektrik/vol32/iss1/1

This Article is brought to you for free and open access by TÜBİTAK Academic Journals. It has been accepted for
inclusion in Turkish Journal of Electrical Engineering and Computer Sciences by an authorized editor of TÜBİTAK
Academic Journals. For more information, please contact [email protected]
 Turkish Journal of Electrical Engineering & Computer Sciences Turk J Elec Eng & Comp Sci
(2024) 32: 1 – 20
http://journals.tubitak.gov.tr/elektrik/
© TÜBİTAK
Research Article doi:10.55730/1300-0632.4052

Traffic signal optimization using multiobjective linear programming for

oversaturated traffic conditions

Mustafa Murat COŞKUN1 , Cevat ŞENER1,2 , İsmail Hakkı TOROSLU1,3∗

1
Department of Computer Engineering, Faculty of Engineering, Middle East Technical University, Ankara, Turkiye

Received: 20.02.2023 • Accepted/Published Online: 28.11.2023 • Final Version: 07.02.2024

Abstract: In this study, we present a framework designed to optimize signals at intersections experiencing oversaturated
traffic conditions, utilizing mixed-integer linear programming (MILP) techniques. The proposed MILP solutions were
developed with different objective functions, namely a reduction in the total remaining queue and fair distribution of
the remaining queue after each signal cycle. Our framework contains two distinct stages. The initial stage applies two
distinct MILP methodologies, while the subsequent stage employs a neighborhood search method to further reduce the
delays associated with the green signal timings derived from the first stage. Ultimately, to evaluate their effectiveness
across various intersections, we employed the HCM 2000 delay model for all the models we developed. Our experimental
results show that the proposed approach reduces the delay significantly for various intersection designs.

Key words: Mixed-integer linear programming, traffic signal optimization, signalized intersections, oversaturated
conditions, deterministic queuing

1. Introduction
In most metropolitan areas, traffic congestion has escalated to significant levels due primarily to population
growth and urbanization, emerging as a pressing issue for both residents and decision-makers [1]. Therefore, the
most important objective of traffic engineering is to create more habitable cities by reducing traffic congestion
within urban traffic networks.
Although there is no universally agreed definition of traffic congestion, it can be described as a situation
that arises when traffic demands on road networks or the volume of vehicles seeking access to these networks
surpass the network’s capacity. In simpler terms, urban traffic encounters congestion when the road networks’
capacity falls below the traffic demands [2].
Traffic signal control strategies are primarily designed to optimize traffic flows at intersections and in
urban road networks with the most efficiency. Researchers have focused on various goals in their efforts to
formulate traffic control strategies for efficient traffic management. These objectives can be in forms such as
minimizing the total number of stops, maximizing traffic flow, reducing total queue lengths, and reducing fuel
consumption. Furthermore, traffic control systems must consider the safety of both vehicles and pedestrians at
intersections.
To develop effective traffic signal control management, a variety of optimization techniques and artificial
intelligence-based algorithms have been employed. An effective traffic signal control system should be cost-
effective, easy to implement, and applicable to real-world conditions. It should aim to maximize traffic flows at
∗ Correspondence: [email protected]

1
This work is licensed under a Creative Commons Attribution 4.0 International License.
 COŞKUN et al./Turk J Elec Eng & Comp Sci

intersections, considering the traffic volumes in each direction. In the present study, we use mixed-integer linear
programming (MILP) approaches because of their clarity and cost efficiency in solving traffic signal control
optimization problems.
Vehicles are required to come to a stop and wait at intersections when they approach while the traffic
signal displays a red light. In instances in which the traffic demand is relatively low (i.e. undersaturated), it is
possible to design the signal cycle in such a way that vehicles that have come to a stop at the preceding red light
can go through the intersection during the next green phase. However, in cases of oversaturation, no matter
how the cycle length is configured, there will still be some vehicles left at the intersection after the green phase
finishes.
If oversaturated traffic conditions are not handled at intersections, they can get worse at each signal cycle.
Therefore, until the traffic flow rate drops below the capacity of the road, the traffic signal cycle settings of the
intersections must be adjusted to reduce the stress of the traffic. The aim of the present work was to reduce
the queue left at intersections after each cycle in order to reduce the delays of vehicles due to oversaturated
conditions.
This paper introduces a novel framework that includes MILP approaches, each with different objective
functions aimed at optimizing fixed-time traffic signal control for isolated signalized intersections facing oversat-
urated conditions. The objective functions were designed to address the limitations of MILP models previously
explored in the literature. They include minimizing the total residual queue length and fairly distributing resid-
ual queues among lane groups under oversaturated conditions. Our models assume a fixed and uniform vehicle
arrival pattern, as seen in linear programming methods in the literature [3, 4]. We evaluate the developed
models based on delay performance criteria and queue length and analyze the impact of different phase designs
on delays.
Within our proposed framework, a new MILP model, abbreviated as MMQLM, was developed to be able
to handle the fairness issue regarding residual queues, which is observed in the MILP model inspired by Liu
(2008). We demonstrate that this new approach produces better solutions in most cases. Additionally, the
framework includes a novel neighborhood search phase to enhance the solution quality obtained from the MILP
models. The main contributions of our paper can be summarized as follows:

• We reformulated the maximization of departure rate formula from Liu (2008) by minimizing the residual
queue, simplifying the model.
• We utilized the HCM 2000 delay model to assess the actual impact of the proposed queue-based solutions
and to evaluate the quality of these solutions.
• We introduced an alternative optimization function that fairly distributes the residual queues across all
lanes.
• We developed an end-to-end framework for implementing the above-mentioned MILP solutions, followed
by a novel neighborhood search algorithm to further enhance the solution’s quality.
• We validated the success of the proposed framework across various intersection architectures using simu-
lation tools.

This paper is organized as follows: Section 2 introduces related work and compares it with our proposed
model. Section 3 introduces and defines the core concepts in traffic engineering. Section 4 describes the details
of the linear programming-based approaches proposed for traffic signal optimization. Section 5 presents the
results of further experiments and, finally, section 6 contains our conclusions.

2
 COŞKUN et al./Turk J Elec Eng & Comp Sci

2. Related work
The main objective of traffic signal optimization is to minimize vehicle delays at intersections. Estimating
delays is not a straightforward task, especially when considering the nonlinearity of the HCM 2000 delay
estimation formula. Consequently, classical linear programming-based optimization methods are not always
directly applicable. To deal with this complexity, a variety of approaches have been employed, including neural
networks [5, 6], fuzzy logic [8, 24], reinforcement learning [9–13], and genetic algorithms [1, 14–17].
Since the signal optimization problem is inherently very complex, linear programming-based methods
should have certain assumptions, like the constant speed of vehicles [3, 4, 18]. In the present work, we made
similar assumptions, and we showed that the method proposed for isolated oversaturated intersections in [4] can
be further improved. In addition to optimizing the signals for isolated intersection, recent research has focused
on optimizing signals of multiple intersections.
The signal optimization for arterial roads is an important variant. The approach used in [4] has been
extended in [19] as a two-stage arterial signal coordination model. [19] presents a two-way signal coordinated
control method aimed at enhancing the traffic efficiency of oversaturated arterial traffic. The approach involves
establishing a traffic flow model based on LWR theory, utilizing vehicle trajectory data for model verification,
and introducing a coordinated control model that optimizes cycle length, green time, and signal coordina-
tion at consecutive intersections through a multiobjective algorithm to maximize throughput and minimize
delay simultaneously.
The complexity of the signal optimization problem increases when attempting to optimize signals at
multiple intersections simultaneously. Most recent work, such as [20] and [21], attacks different versions of this
problem. Due to the complexities of these optimization problem variants, mainly heuristic-based approaches
are employed.
[20] proposes a region-based evaluation particle swarm optimization algorithm, REPSO, with solution
libraries to provide a signal timing scheme for undersaturated and oversaturated traffic flow states in real time
and a region-based evaluation strategy to reduce the number of fitness evaluations.
A two-stage game perimeter control, TSGPC, model is proposed in [21] that integrates macro and micro
control based on game theory to allocate regional regulation quantity to each intersection reasonably and to
optimize signal timing schemes, respectively, where the queue length predicted by the Kalman filter is used as
the basis of games.

3. Traffic engineering
There are various types of intersections and the present study focuses on signalized intersections. In urban
areas, signalized intersections typically fall into three categories: isolated intersections, arterial networks, and
general (grid) networks [22]; herein we focus on isolated intersections.
Traffic signal control systems can be classified into three main categories. Fixed-time traffic signal control
aims to maximize traffic flow at intersections, with signal settings based on historical data and minimal changes
throughout the day. As a result, fixed-time traffic signal control strategies struggle to adapt to fluctuations in
traffic demands. The other two methods, namely traffic-actuated and adaptive traffic control strategies, adjust
signal settings in real time based on detected traffic demands through sensors or technologies like GPS. However,
these methods are costlier than fixed-time strategies.

3
 COŞKUN et al./Turk J Elec Eng & Comp Sci

Traffic flow in a road network can be expressed as the total number of vehicles that pass through a point
in a certain time interval. The total number of vehicles is generally expressed as the hourly traffic volume
(demand) or simply the flow rate. Traffic flows can also be expressed in lanes at signalized intersections.

In traffic engineering, the first step is phase design. Using the structures of phases, the number of vehicles
from different directions that can pass through an intersection is determined. This design process involves
assessing traffic movements and flow directions while considering intersection geometry to prevent potential
conflicts. The number of phases may vary depending on intersection types and geometries.

Traffic signals generally have three signal durations: actual green time, yellow time, and all-red time.
Actual green time allows vehicles with the right of way in a phase to proceed through the intersection.

The traffic signal cycle, also known simply as the cycle, comprises a sequence of signals in which all phases
gain the right of way through the intersection. Once the cycle is completed, the starting phase regains the right
of way. Cycle length refers to the time it takes to complete a cycle.

A lane group consists of one or more lanes, and this information is used for different purposes in traffic
engineering, such as determining the level of service, queuing, capacity, and delay analysis [23]. Among the lane
groups with green time in a phase, the one(s) with the highest flow ratio is called the critical lane group(s).
Since noncritical lane groups have lower traffic demand, their demand is considered satisfied when the demand
of the critical lane group is met. Thus, conditions for critical lane groups are specifically defined.

One critical parameter in effectively designing a traffic signal control system at an intersection is the
saturation flow rate. The Highway Capacity Manual (HCM) defines the saturation flow rate as ”the equivalent
hourly rate at which previously queued vehicles can traverse an intersection approach under prevailing condi-
tions, assuming the green signal is available at all times and no lost times are experienced, in vehicles per hour
or vehicles per hour per lane” [23, p. 61].

The saturation flow rate at intersections is affected by several factors, making it nearly impossible to
obtain an exact value. In many studies, a lane-based saturation flow rate of 1800 veh/h/lane has been assumed
[24? ]. HCM also recommends this value when vehicle speeds are below 50 km/h [23, p. 172]. Hence, in the
present study, we assume a lane-based saturation flow rate of 1800 veh/h/lane.

Given a saturation flow rate ( s ), the effective green time ( g ) of a lane group, and the cycle length ( C ),
the capacity ( c ) of the lane group can be calculated as c = s ∗ (g/C) .

Another important parameter for understanding queue formation at signalized intersections and accu-
rately measuring vehicle delays is the q/c (volume–capacity) ratio. Traffic congestion becomes significant as
the hourly traffic volume ( q ) approaches the lane group’s capacity. Therefore, determining the q/c ratio for
each lane group at intersections is essential before calculating delays and queuing. The q/c ratio is computed
by dividing the hourly traffic volume ( q ) by the lane group’s capacity, yielding the degree of saturation ( Xc )
or the volume-to-capacity ratio: Xc = q/c .

If the degree of saturation of a lane group exceeds 1, assuming a uniform arrival pattern, queues form at
every signal cycle, continuously increasing in size. Conversely, if the degree of saturation of a lane group is less
than or equal to 1, the queue accumulated during the cycle is cleared, leaving no residual queue for the next
cycle. These conditions are known as oversaturated and undersaturated conditions, respectively.

4
 COŞKUN et al./Turk J Elec Eng & Comp Sci

4. Linear programming-based solutions to the traffic signal control problem

The delay calculation formula in the HCM 2000 delay model (given in Appendix B) is nonlinear, as outlined by
HCM [23]. Additionally, for intersections, optimizing the cycle length is very important in minimizing average
control delay. The total cycle length should balance between not too short and not too long, as highlighted by
Papageorgiou [25]. If the cycle length is too short, the effective green times become insufficient to clear the queue
of waiting vehicles. Consequently, the proportion of total lost time in a phase over the total cycle length rises,
potentially leading to an increase in average control delay, which might seem contradictory to expectations.
Conversely, if the cycle length is too long, vehicles in the queue are forced to wait longer for the next green
phase, which can also result in a higher average control delay at the intersection. Therefore, optimizing the
total cycle length can also be a reasonable objective for an optimization program. However, it is essential to
note that this objective introduces nonlinearity into the problem. Further, for oversaturated cases, regardless
of the length of the cycle, there will always be a residual queue left after the effective green time. Thus, there
is a large amount of research focused on optimizing the queue length for fixed-time traffic signal strategies [22].
For these reasons, our paper also focuses on optimizing queuing length using linear programming models.

4.1. Notations and problem specification

Table 1 lists the notations used throughout the paper.

Table 1. Nomenclature.

Notations
P: Set of signal phases indexed by p
N: Number of phases
CG: Set of critical lane groups
Gp : Set of all allowable lane groups in the pth phase
ni : Number of lanes in the lane group i
ωi : Demand ratio of a lane group i
Ω: Total demand ratio of an intersection
R: Conversion coefficient
ai : Allocation ratio of a lane group i
λi : Arrival rate per second for lane group i
θi : Lane-based saturation low rate per second (veh/s/lane) for lane group i
C: Total cycle length
L: Total lost time per cycle
xp : Effective green time for the phase p (decision variable)
p
gmin : Minimum effective green time for the phase p
p
gmax : Maximum effective green time for the phase p

In the present study, three methods were proposed to handle oversaturated conditions in signalized
isolated intersections. These methods, MTQLM, MMQLM, and NSM, which we will introduce shortly, are
specifically intended for applications in which the traffic volume exceeds the capacity, indicating oversaturated
conditions. On the other hand, if the intersection has a higher capacity than traffic, signifying undersaturated
conditions, strategies to minimize the residual queue length are no longer needed. The last method, namely
NSM, is designed to be applied as a follow-up phase of the previous two methods. To assess and compare these
methods, we employed the HCM 2000 delay model to calculate delays corresponding to different green light
settings.

5
 COŞKUN et al./Turk J Elec Eng & Comp Sci

The entire framework is presented in Figure 1. As depicted in the figure, it first gets the parameters of the
intersection and scenario to optimize: lane-based saturation flow rates, hourly traffic volumes, phase design, lost
time, cycle length, yellow time, and all red time. Then it determines the critical lane group(s) as the degree of
the saturation of a critical lane group defines the saturated condition of an intersection. Thus, if Xc is less than
or equal to 1, the algorithm directly stops as the intersection is found to be undersaturated. On the other hand,
once we determine that the intersection and its associated traffic fall into the oversaturated category by having
Xc greater than 1, the two-stage process is followed. Initially, we apply the MTQLM and MMQLM methods,
and then we select the one yielding the superior delay performance. Subsequently, the chosen method’s results
are further enhanced through the application of the NSM method and used as the effective green times for the
intersection.

START

Input
Lane-Based Saturation Flow Rates,
Hourly Traffic Volumes Phase
Design, Lost Time, Cycle Length,
Yellow Time, All Red Time

Determine
Critical Lane Groups

Calculate Xc

Xc ≤ 1 STOP

Stage 1
Apply both MILP Models
in any order

Determine Effective Green Determine Effective Green

Times with MTQLM Times with MMQLM

Select the one with less

1-hour Average Control Delay

Stage 2 Improve the Effective Green

Times obtained in Stage 1
with NSM

STOP

Figure 1. The flow diagram of our framework.

6
 COŞKUN et al./Turk J Elec Eng & Comp Sci

Figure 2. Intersection 1 phases.

To explain the linear programming-based approaches, we designed a four-legged intersection with four
distinct phases, specifying its parameters as presented in Table 2. Given these parameters, with a cycle time ( C )
of 135 s and a total loss time for this cycle ( L ) calculated as 2 (lost time per phase) × 4 (number of phases) +
1 (all − red time per phase) × 4 (number of phases) = 12 s , the total assignable effective green time ( C − L )
amounts to 123 s. In Figure 2, we provide an illustration of a sample intersection. To demonstrate the developed
approaches, we will utilize scenario 1.1, as outlined in Table 2.

Table 2. Parameters (left) and scenario (right) of intersection 1.

Parameters
Initial Cycle Length 135 s
Minimum Effective Green Time per Phase 9s Left (veh/h) Through (veh/h) Right (veh/h)
Lost Time per Phase 2s West 300 1800 144
Yellow Time 3s North 50 350 50
All Red Time per Phase 1s East 156 550 100
Number of Phases 4 South 75 400 75
Number of Lane Groups 6
Preassumed Range Value (δ) for NSM 5

4.2. Solutions for oversaturated scenarios

We present two mixed-integer linear programming (MILP)-based solutions for addressing oversaturated traffic
scenarios. The first approach is from [4] and focuses on minimizing the residual queue length after a traffic cycle
is completed. However, this method, in certain cases, unfairly extends queues in some lane groups, resulting

7
 COŞKUN et al./Turk J Elec Eng & Comp Sci

in increased delays. To overcome this issue, we adapted the approach to minimize the maximum queue length
within lane groups following the cycle. Although this adjustment may lead to a slight increase in the total queue
length, it often results in reduced overall delays for specific scenarios. Additionally, changing the distribution
of green times to various phases can further reduce total delay. Therefore, we introduce a novel neighborhood
search method to identify solutions with improved delay metrics. This approach utilizes solutions obtained from
MILP methods and, instead of conducting an exhaustive search to determine optimal green times, explores the
neighborhood of values derived from the MILP solutions, enhancing their quality. We demonstrate the high
effectiveness of this approach, as it rapidly identifies either the best solution or a solution very close to the
optimal one.

4.2.1. Minimize total queue length method (MTQLM)

MTQLM is a MILP approach designed to minimize the total remaining queue length after each cycle. By
reducing the total remaining queue length, the approach simultaneously maximizes the number of vehicles that
can pass through the intersection, ultimately boosting throughput.
This algorithm is proposed in the work of [4], where the proposed solution assumes for expected over-
saturation of traffic over a defined time period. It aims to determine the optimal traffic light durations for
each phase to maximize total throughput throughout the oversaturated period. In contrast, we operate on the
assumption that we only have knowledge of the traffic volume for the upcoming time period, which is expected
to be oversaturated, and our objective is to determine traffic light durations to minimize the queue length after
a single cycle. Therefore, we modified their approach slightly for this version.
As mentioned previously, when Xc is greater than 1, specific critical lane groups will consistently have a
residual queue after each cycle. This indicates that the effective green time is insufficient to clear these queues
within the effective green period. Consequently, the residual queue continues to grow during the analysis period,
leading to persistent oversaturation in these lane groups.
In this method, while MTQLM attempts to clear the queues for some lane groups, it may leave residual
queues for others. Since MTQLM aims to minimize the total remaining queue length, it allocates the appropriate
effective time to clear queues in different critical lane groups as soon as it finishes clearing a queue in any specific
group. For each critical lane group i within the set of critical lane groups CG ( i ∈ CG ), and the corresponding
phase p within the set of phases p ∈ P , the following constraint applies for oversaturated traffic volume (Total
discharging ≤ Cumulative arriving): xp θi ni ≤ λi C .
Another constraint in this method involves determining minimum effective green times for phases. A
traffic signal setting must ensure pedestrian safety to allow pedestrians to traverse the intersection securely.
Although the present study did not specifically address pedestrians at intersections, minimum effective green
times were still determined for the phases.
For each phase p within the set of phases P, the constraints for minimum effective green times can be
p
expressed as xp ≥ gmin . The final constraint in this approach guarantees that the total assignable effective
green time is equal to the sum of the effective green times for each phase. The total assignable effective green
times are also equal to the difference between the cycle length and the total lost time, represented as follows:

∑
xp = C − L (1)
p∈P

8
 COŞKUN et al./Turk J Elec Eng & Comp Sci

∑ ∑
minimize p∈P (λi C − ni xp θi ) (M inimize T otal Res. Queue)
i∈Gp
subject to
xp θi ni ≤ λi C , ∀i ∈ CG (Discharge Constraint) (2)
p
∑ x p
≥ g min , ∀p ∈ P (M in. Green Constraint)
p∈P x p
= C − L , ∀p ∈ P (Cycle Constraint)

In scenario 1.1, the critical lane groups identified are 1, 2, 3, and 6, while lane groups 4 and 5 are classified
as noncritical in this context. Thus, MTQLM calculates the effective green times for each phase by considering
these critical lane groups and their associated constraints. For this scenario, MTQLM determines effective green
times of 48, 22, 20, and 33 s for phases 1 through 4, respectively.
Using the effective green times generated by MTQLM for scenario 1, after 30 cycles, the total remaining
queue (residual queue) lengths were calculated as 364.5. A detailed breakdown of the residual queue and its
percentages for each lane group is presented in Table 3.
In oversaturated conditions, the HCM 2000 delay analysis gives similar results for both 15-min and 1-h
durations. Thus, we conducted delay analysis for a 1-h duration in these cases. Employing the HCM 2000 delay
analysis for scenario 1.1 over a 1-h analysis period, we determined the intersection’s average control delay as
134.30 s. Additionally, the average control delays for each lane group were calculated as 67.17, 115.00, 99.80,
35.65, 58.53, and 548.39 s, respectively, for lane groups 1 to 6.

Table 3. Residue queue results for scenario 1.1 determined by MTQLM (left) and by MMQLM (right).

Total Residue Residue Queue Total Residue Residue Queue

Lane Groups Total Arriving (veh) Lane Groups Total Arriving (veh)
Queue (veh) Percentage (%) Queue (veh) Percentage (%)
1 2187 27 1 1 2187 342 15
2 337.5 7.5 2 2 337.5 52.5 15
3 506.25 11.4 2 3 506.25 86.4 17
4 731.25 0 0 4 731.25 0 0
5 175.5 0 0 5 175.50 0 0
6 618.75 318.6 51 6 618.75 93.6 15

Analyzing the residual queue results for each lane group in scenario 1.1 reveals that MTQLM leaves
disproportionately long residual queues in certain lane groups. Notably, almost all of the total remaining
queues accumulate in lane group 6. In this scenario, 50% of the vehicles arriving at lane group 6 remain in the
queue and, as the analysis period progresses, these vehicles are forced to wait for one or more green signal cycles
to pass through the intersection. This condition is mainly due to the remarkably high degree of saturation in
lane group 6, which is calculated as 2.06.
Although MTQLM works well in optimizing the total residual queue, it exhibits certain limitations as it
may result in exceptionally lengthy residual queues in specific lane groups. Consequently, this leads to persistent
oversaturation in these lane groups as the analysis period extends, subsequently contributing to significantly
higher average control delays in these lanes when compared to others that experience no residual queues after
each cycle. In simple terms, MTQLM appears to penalize certain lane groups while minimizing the total
remaining queue, thus causing an imbalance in the allocation of effective green times to different phases.

4.2.2. Minimize maximum queue length method (MMQLM)

MMQLM is an alternative MILP approach aimed at addressing the issue of fairness related to residual queue
length and average control delay in the context of the MTQLM approach. Similar to MTQLM, MMQLM

9
 COŞKUN et al./Turk J Elec Eng & Comp Sci

employs the same set of constraints for allocating effective green times to the phases. The key distinction
between MMQLM and MTQLM is in their respective objective functions.
While MTQLM focuses on minimizing the overall residual queue length at the intersection, MMQLM’s
objective is to minimize the maximum remaining queue length in any lane group after each cycle. In other
words, it aims to distribute the remaining total queue among critical lane groups following each cycle. To
achieve a fair allocation of the remaining queue length to each critical lane group, this method employs the
concept of the demand ratio, which differs slightly from the flow ratio. As mentioned previously, the flow ratio
of a critical lane group is determined by dividing the total traffic volume by the total saturation flow rate.
The total saturation flow rate is computed as the sum of the saturation flow rates for each lane within a lane
group. In the present study, the saturation flow rate per lane was assumed to be 1800 vehicles per hour per
lane. Consequently, the total saturation flow rate is equal to the product of the number of lanes in a critical
lane group and the saturation flow rate for a single lane.
On the other hand, the demand ratio for a critical lane group is calculated by dividing the total traffic
volume by the saturation flow rate per lane. The formula for calculating the demand ratio of a critical lane
group is as follows: ωi = qi /si . Once the demand ratios for each critical lane group have been calculated, the
total demand ratio is determined by summing these individual demand ratios. To obtain the total demand ratio
for an intersection, the following formula is used:
∑
Ω= ωi (3)
i∈CG

Next, the allocation ratios for each critical lane group are determined by dividing each individual demand
ratio by the total demand ratio of the intersection, represented as ai = ωi /Ω .
Unlike the MTQLM approach, MMQLM aims to minimize the maximum possible residual queue in any
critical lane group by adjusting its allocation ratio in the residual queue formula. Consequently, when the
saturation flow rate per lane is the same for all lanes, each lane within a critical lane group will have a residual
queue after each cycle with approximately the same percentage. The MILP formulation is illustrated in the
following equation:

minimize maxi ∈ CG ((λi C − ni xp θi )/ai ) (M in. M ax Res. Queue)

subject to
xp θi ni ≤ λi C , ∀i ∈ CG (Discharge Constraint) (4)
p
∑ x ≥p gmin , ∀p ∈ P
p
(M in. Green Constraint)
p∈P x = C − L , ∀p ∈ P (Cycle Constraint)

In scenario 1.1, MMQLM computes the effective green times for each phase as 41, 19, 35, and 28 s,
respectively. Using the effective green times provided by MMQLM, we determined a residual queue analysis
for scenario 1.1 to determine the total queue length after 30 cycles, which has been found as 574.5. A detailed
breakdown of the residual queues and their percentages for each lane group can be found in Table 3.
By employing HCM 2000 delay analysis over a 1-h period, the intersection’s average control delay is
determined to be 127.09 s for scenario 1.1. Additionally, the average control delays for each lane group are
determined as 136.93, 173.64, 168.03, 42.32, 65.29, and 150.68 s for each lane group. In scenario 1.1, MMQLM
generates a total residual queue after 30 cycles of 574.5, which corresponds to a 57.6% increase compared to
MTQLM. However, it is worth noting that MMQLM allocates residual queues to all critical lane groups with
nearly equal percentages. This suggests that the MMQLM approach provides a fairer treatment of critical

10
 COŞKUN et al./Turk J Elec Eng & Comp Sci

lane groups than MTQLM. Furthermore, for scenario 1.1, the MMQLM approach reduces the average vehicle
delay at the intersection by 5.4% compared to MTQLM, resulting in an average delay of 127.09 s per vehicle.
Consequently, it is safe to conclude that, for scenario 1.1, selecting the MMQLM approach is more suitable in
terms of minimizing the average vehicle delay at the intersection and fairly distributing residual queues.
Finally, it is important to note that minimizing the total residual queue at an intersection does not always
correspond to a reduction in the average control delay in certain traffic scenarios. Better delay outcomes are
typically achieved by setting values closer to those that minimize queue lengths. Therefore, we propose a simple
heuristic to search for improved settings that can enhance delay performance.

4.2.3. Neighborhood search method (NSM)

The neighborhood search method (NSM) corresponds to the second stage of our framework for oversaturated
cases, as depicted in Figure 1. In some oversaturated traffic scenarios, the choice between the MTQLM and
MMQLM approaches depends on which one yields better results concerning the HCM 2000 delay model.
Consequently, for a given traffic scenario, these two methods are initially executed in the first stage and compared
based on their average control delay.
During the second stage of the workflow, NSM is employed to search for more optimal effective green
times with a focus on reducing delay. NSM utilizes the effective green times obtained in the first stage as a
starting point to enhance them. This search for improvements is limited to a predefined range, which extends
from −δ to +δ around the best solution from the first stage. The overall flow, incorporating NSM, follows
these steps:

1. Obtain effective green times from the first stage using both MTQLM and MMQLM.

2. Determine the minimum and maximum limits for each effective green time using a predefined range
constant ( δ ).

3. Within the specified limits for each phase, search for an improved green time that reduces the average
control delay based on the HCM 2000 delay formula.

It is important to note that NSM does not guarantee the discovery of the global optimum average control
delay within the predefined range at all times.
In contrast, a simple exhaustive search (brute search) algorithm can potentially identify the global
minimum delay by evaluating all possible combinations of green time settings. However, as the number of
phases and the total cycle length increase for an intersection, the exhaustive search algorithm may require an
extensive amount of time to complete.
For scenario 1.1, NSM enhances the effective green times by exploring within a range of δ (set to 5 s)
around the effective green times obtained in the first stage. This results in new effective green times of 46, 18,
33, and 26 s for phases 1 to 4, with an average control delay of 110.74 s. This represents a 5.18% improvement
compared to the delay obtained by MMQLM, which was 127.09 s.
The number of iterations for NSM can be expressed as (2δ)N , while the number of iterations for the
p
exhaustive search algorithm can be expressed as (gmax )N .
p
For this intersection, gmin is determined as 9 s, which sets the minimum effective green time for a phase.
p
p
The maximum effective green times of a phase are computed as gmax = C − L − (N − 1) ∗ gmin .

11
 COŞKUN et al./Turk J Elec Eng & Comp Sci

As this equation illustrates, the number of iterations for the exhaustive search algorithm increases as the
cycle length and the number of phases increase. On the other hand, NSM has only a small increase in iterations,
dependent solely on the number of phases.
For the present study, both NSM and the exhaustive search algorithm were executed 10 times, and the
average execution time was computed to compare the efficiency of these two algorithms. In the case of an
intersection with 4 phases, a minimum effective green time of 9 s, and a range value δ of 5 s, the mean runtime
for the NSM approach was calculated at 0.13 s, whereas it was 35.34 s for the exhaustive search. NSM was
set to perform 10,000 iterations for this experiment, while the exhaustive search executed more than 84 million
iterations. For this case the NSM algorithm is approximately 270 times faster than the exhaustive search
algorithm, with the number of iterations reflecting their respective execution times.
As mentioned earlier, NSM does not guarantee the discovery of the global minimum delay every time.
When NSM and the exhaustive search algorithms were tested using scenario 1.1, the average control delays
for the intersection were determined as 110.74 s with the NSM method versus 107.53 s with the exhaustive
search method. Although the difference is minimal, NSM’s execution time is significantly faster than that of
the exhaustive algorithm.

5. Experiments
Intersection 1 with four legs was introduced and examined in the previous section, while the methods MTQLM,
MMQLM, and NSM were explained and applied to this sample intersection under a single scenario. This section
will analyze two more intersections with different characteristics using the same three methods: intersection 2
with three legs and intersection 3 with four legs but with more lanes than the first. Here, these two intersections
are tested under 4 and 5 scenarios, respectively. Moreover, the last one was studied with two phase design
alternatives to show their effect on the saturation condition.
All these 45 experiments (3 MILP methods applied to 2 intersections under a total of 5 scenarios plus
1 intersection under 5 scenarios with 2 phase designs) were implemented using Pulp, an open-source Python
library for modeling and solving linear programming problems. It provides a high-level interface for defining
linear programs in which you need to define decision variables you want to optimize, an objective function to
minimize or maximize, and constraints on these decision variables. Its default solver, CBC or COIN-OR Branch
and Cut, was used during these experiments. It is a general-purpose, open-source solver and is available for
various platforms. Appendix A exemplifies the implementation details by applying MTQLM to intersection 1
under scenario 1.1.

5.1. Analyzing intersection 2

Intersection 2, illustrated in Figure 3, is designed as a T-shaped intersection and has a total of 11 lanes. These
lanes form 6 distinct lane groups, depending on the various movements within the intersection. Figure 3 shows
that the phase design includes several instances of repeated lane groups. For instance, lane group 9 is granted
the right of way in more than one phase. In contrast to the first intersection, all movements at this location
are separated from each other, and there are additional lanes allocated for each movement. To assess the linear
programming approaches introduced above, 4 distinct traffic scenarios, labeled 2.1 to 2.4, were generated for
this particular intersection. The parameters and hourly traffic volumes for each lane group in every scenario for
intersection 2 are given in Table 4.

12
 COŞKUN et al./Turk J Elec Eng & Comp Sci

Figure 3. Lane groups (left) and phase design (right) of intersection 2.

Table 4. Parameters (left) and scenarios (right) of intersection 2.

Parameters
Initial Cycle Length 90 s Lane Groups
Minimum Effective Green Time per Phase 8s Scenario No 1 (w-r) 2 (w-s) 3 (e-s) 4 (e-l) 5 (s-r) 6 (s-l)
Lost Time per Phase 2s 2.1 150 1872 990 550 110 990
Yellow Time 3s 2.2 180 1872 1080 600 120 1080
All Red Time 1s 2.3 165 2028 990 600 110 1170
Number of Phases 3 2.4 165 2028 990 700 170 1125
Preassumed Range Value (δ) for NSM 5

MTQLM
MMQLM
Average Control Delay (sec/veh)

500
NSM

400

300

200

100

0
2.1 2.2 2.3 2.4

Scenario No

Figure 4. Delay comparison for intersection 2.

13
 COŞKUN et al./Turk J Elec Eng & Comp Sci

In this intersection, all scenarios produce critical lane groups with a degree of saturation values exceeding
1, resulting in oversaturated conditions. Delays are then compared between MTQLM, MMQLM, and NSM.
The 1-h HCM 2000 average control delay results for these scenarios are depicted in the bar plot in Figure 4.
In particular, NSM consistently produces significantly superior delay results in all scenarios, outperforming
the other two methods. Additionally, MTQLM consistently suffers higher delays than MMQLM and NSM,
particularly as the total hourly traffic volume of the intersection rises. This can be attributed to an increase in
the unbalanced allocation of residual queues in MTQLM as traffic volumes at the intersection increase.
Next, the delay comparison results for the exhaustive search algorithm and NSM are presented in Table 5.
According to these results, NSM managed to identify the global minimum delay in scenarios 2.1, 2.2, and 2.3,
but not in scenario 2.4. Finally, the effective green times and delay found for each scenario are summarized in
Table 6.

Table 5. Delay comparison between NSM and exhaustive search for intersection 2.

Scenario 1-Hour Delay (s/veh) 1-Hour Delay (s/veh)

No NSM Exhaustive Search
2.1 78.88 78.88
2.2 122.55 122.55
2.3 184.97 184.97
2.4 237.62 235.29

Table 6. Effective green times and delays determined for intersection 2.

Scenario 1-Hour Delay

Phase-1 Phase-2 Phase-3
No (s/veh)
2.1 32 24 25 78.88
2.2 31 24 26 122.55
2.3 33 21 27 184.97
2.4 33 24 24 237.62

5.2. Analyzing intersection 3

Intersection 3 is a four-legged intersection, as depicted in Figure 5. It has additional lanes for all left, right, and
straight movements. In particular, all right turns are granted the right of way in two consecutive phases. In
total, intersection 3 contains 14 lanes, which collectively form 12 distinct lane groups corresponding to various
movements.
Unlike intersection 2, two distinct phase designs were created for this intersection to illustrate the impact
of phase design on the intersection’s average control delay. These phase designs, labeled phase design 1 and
phase design 2, along with the movements associated with each phase, are detailed in Figure 6. It is important
to note that even under the same traffic scenario, one phase design may lead to traffic conditions that are
undersaturated, while another may result in oversaturation, as exemplified below.
For this intersection, five different traffic scenarios, denoted as 3.1 to 3.5, were generated. The parameters
and hourly traffic volumes for each lane group in each scenario of intersection 3 are provided in Table 7.
For each scenario at this intersection, the delays experienced by each method were thoroughly examined
and compared. Subsequently, the effective green times and delays identified for phase design 1 & 2 are presented

14
 COŞKUN et al./Turk J Elec Eng & Comp Sci

in Table 8. Here ”n/a” in the Delay column indicates that the corresponding scenario and phase design produce
an unsaturated condition and, hence, the models were not applied.

Figure 5. Intersection 3 - lane groups.

Figure 6. Intersection 3 - phase design-1 (left) and phase design-2 (right).

15
 COŞKUN et al./Turk J Elec Eng & Comp Sci

Table 7. Parameters (left) and scenarios (right) of intersection 3.

Parameters
Lane Groups
Initial Cycle Length 135 s
Scenario No 1 2 3 4 5 6 7 8 9 10 11 12
Minimum Effective Green Time Per Phase 9s
3.1 110 550 165 165 495 275 138 1040 385 110 165 165
Lost Time Per Phase 2s
3.2 110 550 165 165 585 288 138 1040 420 110 165 165
Yellow Time 3s
3.3 130 650 195 195 585 325 163 1040 455 130 195 195
All Red Time 1s
3.4 130 650 195 195 585 325 163 1040 455 130 300 195
Number of Phases 4
3.5 130 720 195 195 585 325 163 1040 455 130 300 195
Preassumed Range Value (δ) for NSM 5

Table 8. Effective green times and delays determined for phase design 1 (left) and 2 (right) of intersection 3.

Scenario 1-Hour Delay Scenario 1-Hour Delay

Phase-1 Phase-2 Phase-3 Phase-4 Phase-1 Phase-2 Phase-3 Phase-4
No (s/veh) No (s/veh)
3.1 28 35 40 20 99.37 3.1 24 41 44 14 n/a
3.2 27 38 39 19 163.14 3.2 22 45 42 14 n/a
3.3 28 37 38 20 193.96 3.3 25 44 39 15 n/a
3.4 28 37 38 20 190.46 3.4 24 40 39 20 125.90
3.5 28 37 38 20 190.72 3.5 27 38 39 19 142.36

As depicted in Figure 7, the delays attributed to phase design 1 and phase design 2 can exhibit significant
differences, as apparent in scenarios 3.4 and 3.5. Moreover, phase design can have the effect of changing the traffic
conditions to undersaturated. Specifically, for this intersection, selecting phase design 2 results in scenarios 3.1,
3.2, and 3.3 being categorized as undersaturated, consequently leading to unreported delay values.

200 Phase Design 1

Phase Design 2
Average Control Delay (sec/veh)

150

100

50

0
3.1 3.2 3.3 3.4 3.5

Scenario No

Figure 7. Delay comparison between phase design 1 and 2 of intersection 3.

5.3. Delay comparison with PTV VISSIM

PTV VISSIM is behavior-based and commercial microscopic simulation software that analyzes and optimizes
traffic flows and tests real-life models. It provides detailed reports (vehicle delay, fuel consumption, average
queue length, etc.) by analyzing models and traffic flows and can visualize models in detail [26].
In this section, intersection 3, previously evaluated with the HCM 2000 delay calculation, will be tested
in the PTV VISSIM simulation environment. The simulation aims to validate our framework by obtaining delay
results comparable to those obtained by the analytical method HCM 2000.
PTV VISSIM has a stochastic nature and uses the Poisson distribution when vehicles enter the network.
Therefore, it can be predicted that there will be a certain difference between the average vehicle delay measured
in PTV VISSIM and the delay obtained analytically with HCM 2000. Furthermore, saturation flow rates cannot

16
 COŞKUN et al./Turk J Elec Eng & Comp Sci

be directly entered as a value for simulation in PTV VISSIM. Instead, they are determined by driver behavior
and car-following logic. In the present study, we attempted to approach the predetermined lane-based saturation
flow rate of 1800 veh/h/lane by adjusting driver behavior parameters. However, it should be noted that it is
nearly impossible to achieve an exact saturation flow rate of 1800 veh/h/lane in the simulation.
In order to simulate intersection 3 and its scenarios, the lane group, their signal heads, and a four-phase
signal controller were created, and all the necessary parameters were supplied. The 3D view of this simulation
setup can be seen in Figure 8. This PTV VISSIM simulation was conducted 10 times for each scenario, with
different random seed values, and each time for an hour. The results of the 1-h control delay of HCM 2000 and
the average vehicle delay of PTV VISSIM can be found in Table 9. Although the analytical and simulation
delays are not the same, this is a predicted result, as explained above. However, the delay variances are minor
except for scenario 3.2. Based on these findings, it can be concluded that the simulations confirm our analytical
results to some extent.

Figure 8. 3D view of intersection 3 simulation in PTV VISSIM.

Table 9. HCM 2000 delay vs. VISSIM delay for each scenario of intersection 3.

1-Hour HCM 2000 1-Hour PTV VISSIM Delay Difference

Scenario No
Average Vehicle Delay (s/veh) Average Vehicle Delay (s/veh) Percentage (%)
1 61.69 64.54 +%4.6
2 71.44 78.53 +%9.9
3 92.54 95.26 +%2.9
4 125.90 129.12 +%2.6
5 142.36 141.96 -%0.3

5.4. Discussion of experiment results

In our experiments, we designed various signalized intersection configurations and different phase designs to
demonstrate the flexibility of our framework, regardless of the intersection type. Furthermore, we tested our
approach across a range of oversaturated scenarios. Our observations revealed that, in most instances, our
optimization approach, specifically aimed at minimizing the maximum queue length, outperformed a simplistic

17
 COŞKUN et al./Turk J Elec Eng & Comp Sci

approach of total queue length minimization.

While MILP models are oriented towards optimizing queue lengths, the actual cost to vehicles is the delay
suffered while waiting at the intersection. Consequently, there is potential for further improvement by refining
signal timings. To address this, we applied our novel neighborhood search heuristic to the solutions derived
from MILP. Notably, we found that in almost all cases these proposed solutions could be further improved with
this more cost-effective heuristic.

6. Conclusions
Isolated signalized intersections are very common on roads. Designing roads to accommodate the demands
of rush hour traffic is often impractical. Consequently, when the traffic demand exceeds the capacity of these
intersections, oversaturation becomes inevitable. This study presents a framework that includes various objective
functions for addressing the traffic signal control problem through the application of linear programming
optimization techniques. Our approach is specifically designed for determining signal settings for oversaturated
conditions for isolated signalized intersections.
We employ a fixed-time signal control strategy with the objective of optimizing traffic control at isolated
signalized intersections. While traffic optimization is often framed as the maximization of traffic flow through
the intersection, we choose to model our problem in terms of the residual queue left after each traffic cycle.
Since, for individual vehicles, the actual impact of any traffic issue is the delay experienced by the vehicle, the
quality of our solution is best assessed by considering the average delay that a vehicle experiences when passing
through the intersection. This is why we utilize the HCM 2000 delay model to evaluate the MILP models that
we explored.
In oversaturation situations, the MTQLM approach, aimed at minimizing overall queue lengths, can lead
to uneven queues in different lane groups. This problem becomes more obvious, particularly in high-traffic
scenarios. As a response, a new method, MMQLM, has been developed. The aim of MMQLM is to achieve
a fairer treatment of lane groups by considering vehicle arrival rates when distributing the remaining queue
length.
Furthermore, it has been observed that the NSM approach, employed to further reduce average vehicle
delays at intersections using the green times derived from the initial stage, yields even lower delay results for
scenarios characterized by oversaturated conditions. Especially since NSM operates as a neighborhood search
heuristic, it works with shorter execution times in comparison to exhaustive search methods. Consequently, this
two-step approach proposed for addressing oversaturated conditions can serve as a feasible alternative to other
linear programming models featured in the literature.
Our experiments with various intersection architectures show that our approach is quite effective. We
also verified the models through a simulation tool.

References

[1] Singh L, Tripathi S, Arora H. Time optimization for traffic signal control using genetic algorithm. International
Journal of Recent Trends in Engineering 2009; 2 (2): 4.
[2] Weisbrod G, Vary D, Treyz G. Economic Implications of Congestion. Technical report, NCHRP REPORT 463,
National Cooperative Highway Research Program, Transportation Research Board, 2001.
[3] Srisurin P, Singh A. Optimal signal plan for minimizing queue lengths at a congested intersection. International
Journal of Traffic and Transport Engineering 2017; 6 (3): 53-63.

18
 COŞKUN et al./Turk J Elec Eng & Comp Sci

[4] Liu H, Balke KN, Lin WH. A reverse causal-effect modeling approach for signal control of an oversaturated
intersection. Transportation Research Part C: Emerging Technology 2008; 16 (6): 742-754.
[5] Srinivasan D, Choy MC, Cheu RL. Neural networks for real-time traffic signal control. IEEE Transactions on
Intelligent Transportation Systems 2006; 7 (3): 261-272.
[6] Castro GB, Hirakawa AR, Martini JSC. Adaptive traffic signal control based on bio-neural network. Procedia
Computer Science 2017; 109: 1182-1187.
[7] Trabia MB, Kaseko MS, Ande M. A two-stage fuzzy logic controller for traffic signals. Transportation Research
Part C: Emerging Technology 1999; 7 (6): 353-367.
[8] Chiu S. Adaptive traffic signal control using fuzzy logic. In: Proceedings of the IEEE Intelligent Vehicles Symposium;
Westlake Village, CA, USA; 1992. pp. 98-1007.
[9] Zeng J, Hu J, Zhang Y. Adaptive traffic signal control with deep recurrent Q-learning. In: IEEE Intelligent Vehicles
Symposium (IV), Proceedings; Changshu, China; 2018. pp. 1215-1220.
[10] Liang X, Du X, Wang G, Han Z. Deep reinforcement learning for traffic light control in vehicular networks. IEEE
Transactions on Vehicular Technology 2018; 1-11.
[11] Shoufeng L, Ximin L, Shiqiang D. Q-learning for adaptive traffic signal control based on delay minimization strategy.
In: IEEE International Conference on Networking, Sensing and Control; Sanya, China; 2008. pp. 687-691.
[12] Van der Pol E, Oliehoek FA. Coordinated deep reinforcement learners for traffic light control. In: 30th Conference
on Neural Information Processing Systems; Barcelona, Spain; 2016.
[13] Xiao N, Yu L, Yu J, Chen P, Liu Y. A cold-start-free reinforcement learning approach for traffic signal control.
Journal of Intelligent Transportation Systems 2021; 26 (4): 1-10.
[14] Park B, Messer CJ, Urbanik T II. Traffic signal optimization program for oversaturated conditions: genetic algorithm
approach. Transportation Research Record 1999; 1683 (1): 133-142.
[15] Chin YK, Yong KC, Bolong N, Yang SS, Teo KTK. Multiple intersections traffic signal timing optimization with
genetic algorithm. In: Proceedings of International Conference on Control System, Computing and Engineering,
ICCSCE 2011; Penang, Malaysia; 2011. pp. 454-459.
[16] Stevanovic A, Martin PT, Stevanovic J. VisSim-based genetic algorithm optimization of signal timings. Transporta-
tion Research Record 2007; 2035 (1): 59-68.
[17] Lee J, Abdulhai B, Shalaby A, Chung EH. Real-time optimization for adaptive traffic signal control using genetic
algorithms. Journal of Intelligent Transportation Systems 2005; 9 (3): 111-122.
[18] Coll P, Factorovich P, Loiseau I, Gomez R. A linear programming approach for adaptive synchronization of traffic
signals. International Transactions in Operational Research 2013; 20 (5): 667-679.
[19] Wang H, Peng X. Coordinated control model for oversaturated arterial intersections. IEEE Transactions on Intel-
ligent Transportation Systems 2022; 23 (12): 24157-24175.
[20] Zhang C, Li JY, Chen CH, Li Y, Zhan ZH. Region-based evaluation particle swarm optimization with dual
solution libraries for real-time traffic signal timing optimization. In: Proceedings of the Genetic and Evolutionary
Computation Conference, GECCO ’23; 2023. New York, NY, USA: Association for Computing Machinery. p. 111-
118.
[21] Li S, Kang L, Huang H, Liu L. A perimeter control model of urban road network based on cooperative-
noncooperative two-stage game. Physica A: Statistical Mechanics and its Applications 2023; 626: 129081.
[22] Eom M, Kim BI. The traffic signal control problem for intersections: a review. European Transport Research Review
2020; 12 (1): 50.
[23] HCM 2000 Highway Capacity Manual. Technical Report. Washington DC, USA: Transport Research Board, 2000.
[24] Shao CQ, Rong J, Liu XM. Study on the saturation flow rate and its influence factors at signalized intersections in
China. Procedia - Social and Behavioral Sciences 2011; 16: 504-514.

19
 COŞKUN et al./Turk J Elec Eng & Comp Sci

[25] Papageorgiou M, Diakaki C, Dinopoulou V, Kotsialos A, Wang Y. Review of road traffic control strategies.
Proceedings of the IEEE 2003; 91 (12): 2043-2065.
[26] Fellendorf M, Vortisch P. Microscopic traffic flow simulator VISSIM. In: Barceló J (editor). Fundamentals of Traffic
Simulation. International Series in Operations Research & Management Science, vol 145. New York, NY, USA:
Springer, 2010, pp. 63-93.

20
 COŞKUN et al./Turk J Elec Eng & Comp Sci

Appendix A: Applying MTQLM to scenario 1.1

To apply MTQLM to intersection 1 under scenario 1.1, we need to calculate the parameters in the method
formulation.

First, based on Figure 2 and Table 3, the per-second arrival rate for each lane group is calculated as
follows, abbreviating Arrival Rate to AR:

λ1 = ARWest-to-Through + ARWest-to-Right = 1800 + 144 (per h) = 0.54 (per s)

λ2 = ARWest-to-Left = 300 (per h) = 0.083 (per s)
λ3 = ARNorth-to-Right + ARNorth-to-Through + ARNorth-to-Left = 50 + 350 + 50 (per h) = 0.125 (per s)
λ4 = AREast-to-Right + AREast-to-Through = 100 + 550 (per h) = 0.18 (per s)
λ5 = AREast-to-Left = 156 (per h) = 0.043 (per s)
λ6 = ARSouth-to-Right + ARSouth-to-Through + ARSouth-to-Left = 75 + 400 + 75 (per h) = 0.152 (per s)

The number of lanes in each lane group is apparent in Figure 2:

n1 = 3, n2 = 1, n3 = 1, n4 = 2, n5 = 1, n6 = 1

As mentioned in Section 2, the saturation flow rate is assumed to be 1800 veh/h/lane; hence:

θ = 1800 veh/h = 0.5 veh/s for each lane

Moreover, Section 3.1 and Table 2 contain the following parameters:

Total cycle length, C = 135 s

Total lost time per cycle, L = 12 s
Minimum effective green time per phase, gmin = 9 s

Lastly, the decision variables can be expressed as below:

Xp : Effective green time for the phase p

Based on the parameters explained above, the MTQLM code in Python and its output are given in
Figure 9 and Figure 10, respectively.

1
 COŞKUN et al./Turk J Elec Eng & Comp Sci

Figure 9. MTQLM code in Python for scenario 1.1.

Figure 10. Output of the MTQLM code for scenario 1.1.

2
 COŞKUN et al./Turk J Elec Eng & Comp Sci

Appendix B: HCM 2000 analytical delay model

According to the HCM 2000 delay model [23, p. 317], the average control delay in a lane group can be calculated
by the following equations:

d = d1 (PF) + d2 + d3
g
0.5C(1 − )2
d1 = C
g (5)
1 − [min(1, X) ]
√ C
8kIX
d2 = 900T[ (X − 1) + (X − 1)2 + ]
cT

Here
d = control delay (s/veh)
d1 = uniform delay (s/veh)
d2 = incremental delay (s/veh)
d3 = initial queue delay (s/veh) (is chosen as 0 assuming no initial queue)
P F = progression adjustment factor (is chosen as 1 for isolated intersections)
X = volume-to-capacity ratio (q/c or q/c) or degree of saturation for lane group
C = cycle length (s)
c = capacity of the lane group (s)
g = effective green time for the lane group (s)
T = duration of analysis period (h)
k = incremental delay adjustment for actuated control (is chosen as 0.5 for fixed-time signal designs)
I = incremental delay adjustment for filtering and metering by upstream signals (1 for isolated intersections)

We studied isolated signalized intersections, which are at least 1.6 km from any other upstream
signalized intersections [23]. For this reason, I , the upstream filtering and metering adjustment factor, is
included in the formula as 1, as HCM suggested. We also studied optimizing fixed-time signal design in the
present study. Therefore, the value k in the formula is taken as 0.5 for the fixed-time signal control. The
progression adjustment factor is calculated and included to explain the effect of coordinated traffic signal
control. In isolated intersections, this value is taken as 1. Finally, we assume no initial queue is formed before
the analysis period, so the control delay d3 was taken as 0 for all computations.
According to the HCM 2000 delay model, in cases in which the degree of saturation is relatively low, i.e.
not too close to 1, the delay is usually caused by the uniform delay d1 part of the formula. In such cases, the
effect of the incremental delay d2 is low as well. However, in cases in which the degree of saturation is very
close to 1 or greater than 1, the effect of d2 is relatively high, as well as the uniform delay. Moreover, if there
is persistent saturation in a lane group, d2 will continue to increase as the analysis period gets longer. It is
crucial to know this relationship to understand the effect of delays while analyzing intersection performance.