SmartLight: A Decentralized MPC-based

Multi-intersection Traffic Signal Optimization Algorithm

Eurick Evardone, Myles Arvin Manalangsang, Victor Jacob Manlapaz, Rodeilyn Jean Omanio​
Electrical and Electronics Engineering Institute, University of the Philippines- Diliman​
<eurick.evardone><myles.arvin.manalangsang><victor.jacob.manlapaz><rodeilyn.jean.omanio>@eee.upd.edu.ph

Abstract—Heavy vehicular traffic congestion takes a heavy factors of congestion in major roads [4]. Management of
toll on the Philippines, negatively affecting the travel time of traffic, specifically in intersections, is improved by adequate
Filipino commuters. One of the primary causes of traffic traffic signal control [5].
congestion along roads are the fixed-time traffic signaling
systems that are widely used in the country. This study focuses To tackle the issues pinpointed by the aforementioned case
on Katipunan-B. Gonzales-Thornton Drive intersection, and studies, dynamic traffic signal control can be achieved using
the Katipunan-University Road-F. dela Rosa intersection, two either a centralized or decentralized approach.[6]. In a
consecutive intersections along Katipunan. Decentralized centralized control approach, there would be one agent to
Model Predictive Control (MPC) algorithms using the control the whole network of intersections. This was done
store-and-forward model and the cell transmission model with
with an MPC-based approach, which has resulted in such a
(i) queue time, (ii) queue length, and (iii) vehicular flow rate
performance parameters are compared to those of the significant improvement in heavy traffic that an MPC
standardized fixed time signaling system. The MPC for the approach was described as “fitting perfectly” [7]. However,
store-and-forward model is generated from a Gaussian Process it has been noted that such a system is difficult to upscale
Regression (Bayesian Optimization) which seeks the green time [8]. An alternative approach would be a decentralized
that produces the best results of the three parameters control approach where each agent controls a single
mentioned. The cell transmission model follows an actuated intersection and shares relevant information with adjacent
signal control based on the aggregated delay time produced by intersections. For this approach, each agent must be trained
predicted cell parameters of the next time step. The centralized with data specific to its assigned intersection. Despite the
models for both the SF model and CTM produced better
advantages of centralized MPC, real-life implementations
performance than the fixed time, as well as servicing more cars
than the standard fixed-time TSC. The decentralized models are difficult to do due to the high computational complexity
however, struggled with the queue lengths and queue times due and communication needed between controllers.
to the asynchronous control between the two close This project implements a dynamic control of traffic
intersections. signal lights using various approaches that are tested using
Keywords—Traffic control, Model Predictive Control (MPC) Simulation of Urban Mobility (SUMO). Two consecutive
algorithm, Store and forward model (SF), Cell transmission intersections along Katipunan Avenue, a major road on the
model (CTM), SUMO (traffic simulation tool). eastern side of the University of the Philippines Diliman
I.​ INTRODUCTION campus, are used as a control setting for this project. The
two intersections are the Katipunan-B. Gonzales-Thornton
Traffic congestion has been one of the biggest problems Drive intersection (in front of one of the gates of Miriam
in the Philippines, especially in Metro Manila. In 2022, the College) and the Katipunan-University Road-F. Dela Rosa
TomTom index ranked the city of Manila as the 9th worst intersection (where Gate 3 of the Ateneo de Manila
traffic-congested city in the world, and 2nd across East and University is located).
Southeast Asia [1]. Traffic congestion not only affects the
travel times of the citizens. It also has implications on the II.​ MPC ALGORITHM
country’s economy, environment, and different societal A.​ Store-and-Forward Model
operations as the traffic in Metro Manila has translated to a For this project, the primary model used was the Store
societal cost of around 20 billion USD annually [2]. and Forward Model. In the context of car traffic, it is a
Several schemes have been used to address the problem traffic flow theory that describes the behavior of vehicles
of traffic in the Philippine setting. In 1977, the Metro Manila moving through a roadway network.
Traffic Engineering and Management Project (TEAM) In this model, vehicles encounter various conditions such
commenced, aiming to address the increasing traffic as traffic signals, congestion, or bottlenecks along their
congestion in the metropolitan area. Numerous traffic light route. When vehicles approach a congested area or a
systems, traffic signs, and markings were installed all over bottleneck, they may come to a temporary stop or slow
the city. Since then, several coding schemes, traffic signal down, creating a queue or "store" of vehicles. As the traffic
control(TSC) systems, and flow management schemes have conditions ease or as vehicles move forward due to signal
been formulated and implemented but the problem of traffic changes or improved roadway conditions, the stored
congestion is yet to be solved [3]. Studies on the analysis of vehicles are then able to proceed along their intended path.
traffic in Metro Manila have identified signalized
intersections as major conflict points concerning traffic Equation (1) shows the standard mathematical model for
demand. To ensure a better vehicular flow and a the Store and Forward Model, which shows how many
minimization of congestion, efficient management of traffic vehicles will be in the store after a given time period.
in intersections is needed. Case studies by the Japan 𝑜𝑢𝑡 𝑆𝑖,𝑗
𝑥(𝑘 + 1) = 𝑥(𝑘) + 𝑇( ∑ τ𝑖,𝑗 · 𝑞𝑖,𝑗 (𝑘) − · ∑ 𝑢𝑖,𝑗(𝑘))​ (1)
International Cooperation Agency (JICA) Project team 𝑖,𝑗ϵ𝐿
𝐶
𝑖,𝑗ϵ𝐿
identified inadequate signal timing, and unreliable offset
In order to adopt this model for optimization purposes,
timing resulting in a lack of signal coordination as key
exit nodes were created, to effectively measure and

©2024 IEEE
maximize flow rate. The optimization code itself functions III.​ METHODOLOGY
by obtaining traffic data and phase data from TraCI, and
A.​ Data Gathering & Manipulation Phase
using that to evaluate Equation (1). Fig 1 shows a
rudimentary summary of the code. 1)​ Traffic Data Collection
Traffic demand data was collected by communicating
with the Metropolitan Manila Development Authority
(MMDA). The data acquired in this activity were only the
AM and PM peak hour traffic volume data per vehicle type
for both Katipunan-B. Gonzales-Thornton Drive
intersection(KBT), and the Katipunan-University Road-F.
dela Rosa intersection(KUF). Thus, the obtained data have
undergone averaging to be distributed into hourly traffic
volume data. Routing diagrams are also present to show the
general traffic flow in the intersection.
Fig. 1. MPC Algorithm
2)​ Data Averaging
B.​ Cell Transmission Model The acquired data from the MMDA contained
The secondary model that will be used is the Cell independent data for the two intersections. The
Transmission Model (CTM), which presents a practical Katipunan-University Road-F. Dela Rosa intersection had
framework for traffic signal control, notably employed in traffic volume data on August 22, 2023 (Tuesday), from 6
urban road network studies. It's a discrete representation am to 8 pm. The morning peak was from 7 am to 8 am while
derived from the Lighthill-Whitham-Richards (LWR) the afternoon peak was from 5 pm to 6 pm. The
model, segmenting the road network into cells, each Katipunan-B. Gonzales-Thornton Drive intersection on the
spanning the distance covered by uncongested traffic within other hand had traffic volume data on February 28, 2023
a single time interval. The CTM's dynamics are captured by (Tuesday), also from 6 am to 8 pm. The morning peak was
the equations in (2) and (3). from 8 am to 9 am while the evening peak was from 6 pm to
Equation (2) closely resembles the store-and-forward 7 pm.
model, illustrating the variation in vehicle count (n). In Due to the discrepancy in data collection, the actual
contrast to road links, the CTM tracks changes in vehicle demand data for the two intersections had different inflow
count within a cell, relying on the inflow of vehicles (f) from and outflow volumes. The study involves a road segment
the current cell (i) and its subsequent cell (i+1). linking the southern end of the Katipunan-B.
Conversely, the second equation defines the vehicle Gonzales-Thornton Drive and the northern end of the
inflow, computed as the minimum among three factors: (i) Katipunan-University Road-F. Dela Rosa. Maintaining equal
the vehicle count in the upstream cell (i−1), (ii) the inflow traffic volumes at both ends is crucial to prevent cars from
capacity (Q) of the current cell, and (iii) a calculation appearing and disappearing mid-route.
involving the ratio of free-flow speed (w) to backward
𝑇𝑜𝑡𝑎𝑙 𝑉𝑜𝑙𝑢𝑚𝑒𝐾𝑈𝐹, 𝑆𝐵
propagation speed (v), multiplied by the difference between 𝑇𝑜𝑡𝑎𝑙 𝑉𝑜𝑙𝑢𝑚𝑒𝐾𝐵𝑇, 𝑆𝐵
𝑖 𝑖
× 𝐿𝑎𝑛𝑒 𝑉𝑜𝑙𝑢𝑚𝑒 = 𝑁𝑒𝑤 𝑉𝑜𝑙𝑢𝑚𝑒 𝐾𝐵𝑇, 𝑆𝐵​ (1)
the holding capacity (N) and the vehicle count (n) within the 𝑇𝑜𝑡𝑎𝑙 𝑉𝑜𝑙𝑢𝑚𝑒𝐾𝐵𝑇, 𝑁𝐵−50 𝑗 𝑗
current cell. 𝑇𝑜𝑡𝑎𝑙 𝑉𝑜𝑙𝑢𝑚𝑒𝐾𝑈𝐹, 𝑁𝐵−50
× 𝐿𝑎𝑛𝑒 𝑉𝑜𝑙𝑢𝑚𝑒 = 𝑁𝑒𝑤 𝑉𝑜𝑙𝑢𝑚𝑒 𝐾𝑈𝐹, 𝑁𝐵 (2)
𝑛𝑖(𝑘 + 1) = 𝑛𝑖(𝑘) + 𝑓𝑖(𝑘) − 𝑓𝑖+1(𝑘) (2)
The two equations are used to transform the volume of
the intersection where the vehicles will come from, as such,
{
𝑓𝑖(𝑘) = 𝑚𝑖𝑛 𝑛𝑖−1(𝑘), 𝑄𝑖(𝑘),
𝑤
𝑣 [𝑁𝑖(𝑘)− 𝑛𝑖(𝑘)] } (3)
the generating edges will then follow the averaging as well.
Contrary to the store-and-forward model, green time and The superscripts i and j are not exponents in the equation,
cycle time are not implicit. Instead, what CTM constitutes these denote the individual flows that enter the road segment
are delay times that are triggered by individual cell being transformed. It is also important to note that 50 is
parameters in the system. CTM defines delay as the subtracted in the northbound equation due to the presence of
additional time beyond free-flow time that a vehicle stays in a U-turn that then transfers cars from the southbound lane to
a cell [9]. the northbound lane of the road segment. This is done to
𝑑𝑖(𝑘) = 𝑛𝑖(𝑘) − 𝑓𝑖+1(𝑘) ​ ​ (4) avoid the averaging done in these vehicles as the ratio is
different for the southbound and northbound lanes. The
Equation (4) shows the delay time computation for each
averaged traffic volume data will then be used as input
individual cell (i), which contributes a delay whenever the
parameters for traffic demand data in SUMO.
outflow of cell (i) at time-step (t) is less than the current cell
occupancy. The total delay time is the sum of all aggregated 3)​ Ocular Inspection
delay times of each cell. An ocular inspection of the two intersections was made
The decision function compares the delay time for the to produce the phase diagram and measure the cycle time of
next time step of the current active phase when stopped to the intersection. The researchers obtained simultaneous
the succeeding phases while taking into account a minimum cycle timings using a stopwatch. Three samples of time
set time for green time. Also, it is important to set a measurements are recorded to take into account deviations
maximum threshold to avoid biasing the whole intersection and human errors that may cause inconsistencies in the
into a single phase only. readings. Fig. 2 shows the phase diagrams for the KBT and
the KUF intersections respectively.

©2024 IEEE
 A summarized version of the cycle timings is shown in generate an XML file consisting of the different flow
tabular form below. Note that a slight difference from other instances. For this project, a total of 41 routes were
conventional traffic light systems, the traffic lights in the generated to accommodate the overlapping routes of the 2
two intersections possess a delay time between the ending of intersections
one phase’s yellow light to the beginning of the next phase’s
3)​ Demand Generation
green light (a time elapses where all the traffic lights in the
Working with the averaged data for the 4-hour period, the
intersection are red). The whole time when another phase is
vehicular data was distributed on hourly periods to simplify
happening at a given moment, all other phases are assumed
the setting up of the flow of vehicles inside the simulation
to be in red light (not active).
and to make the traffic as realistic as possible. To distribute
this 4-hour data, the hourly data distribution for the 4-hour
period from a previously similarly conducted project by
(Magueflor et al.) was used.
In developing the volume distribution code, roads were
classified under three types: source, intermediate, and sink.
Intermediate roads are in between source and sink roads. For
routes that have a direct connection to their destination,
volume generation was straightforward but for routes that
had intermediate roads, vehicles from the source roads were
redistributed across the different possible routes in line with
Fig. 2. Phase Diagram for KBT(left) and KUF(right) intersections MMDA data.
TABLE I. Timer Settings of the Current Fixed-time Traffic Control
As the simulation tool takes in demand through an XML
Phase Green Time Yellow Time Delay Time (all red) file, an XML entry generator was developed through
Katipunan-B. Gonzales-Thornton Drive intersection
Phase 1 (Green) 120s 3s 4s
Python. Each XML entry signifies the number of vehicles
Phase 2 (Red) 40s 3s 4s passing through a route inside the simulation. The
Phase 3 (Blue) 42s 3s 4s highlighted line below shows an example of an XML entry
Phase 4 (Yellow) 26s 3s 4s used for demand generation. The number pertains to the
Katipunan-University Road-F. dela Rosa intersection number of vehicles to be distributed equally over the
Phase 1 (Green) 120s 3s 4s
Phase 2 (Yellow) 40s 3s 4s
beginning and ending time of the flow.
Phase 3 (Red) 42s 3s 4s <flow id ="f_99" type = "phMotor" begin="3600.00" route="kuf-10c" end="7200.00"
number="344" departLane="free" departSpeed="max"/>
Phase 4 (Blue) 26s 3s 4s
This flow definition with vehicle type and route is then
4)​ Vehicular Profiling
fed into a program that generates individual vehicles namely
Vehicles were profiled into two main types: private duaIterate.py. Using the program, certain vehicle attributes
vehicles and motorcycles, to adhere to the two main types of are set to mirror the real driver behavior i.e., departLane of
vehicles along the intersections. A special profile for “free” was selected to mimic vehicle behavior of taking free
motorcycles was created due to their volume, dimensions, lanes in the road, and then switching to the appropriate lanes
and driving behavior. To simplify distribution, a 2:3 ratio of for their corresponding destination.
private vehicles to motorcycles was used. This ratio was
arrived at by sampling from different profiled vehicular 4)​ Vehicle Type Settings
volume data. To make the simulation more accurate with its real-life
counterpart certain vehicle behaviors were set such as the
B.​ Simulation Development Phase
setting of sigma that models imperfections in driver
This phase proceeded simultaneously with the data
behavior. For motorcycles, it should be noted that it was set
gathering phase.
to a lower value (i.e. closer to ideal behavior) as they move
1)​ Road Network Design with quicker reaction times, as compared to regular vehicles.
With SUMO as the chosen traffic simulator, an initial Sublanes were also introduced inside the network to mimic
traffic network was made using SUMO’s integration with motorcycle behavior on the road here in the Philippines.
OpenStreetMap (OSM). OSM, with its database, was used With the sublanes implemented the lateral alignment of the
to create an electronic map of the road network, from this, cars is set to “center” to retain the actual behavior in real
an edge type file was compiled, and through SUMO’s life. Additionally, the dimensions of the vehicles are
integration with OSM, a SUMO Network File was modified to emulate the proportions of real-life vehicles.
generated. Shown in the left of Fig. 3 is the selection of the C.​ Testing Phase
area to be converted into a SUMO network file through
OSM. 1)​ Bayesian Optimization
Using Bayesian Optimization for the Store and Forward
2)​ Route Setup Model, the model maximizes the flow rate of any given
MMDA data was then used to procure the routing phase by tuning the hyperparameters of the system.
scheme of the vehicles. Fig. 7 below was used as a reference Furthermore, in order to prevent the algorithm from
in mapping the routes underperforming in systems where only a few of the roads
In preparation for simulations through vehicle demand have cars, a small incentive that scales with time is added to
generation, the routes that different vehicles take have to be the flow rate to increase the time given.
set-up. Route mode in netedit of SUMO was used to

©2024 IEEE
 2)​ Performance Indicators 𝑆𝑄𝑉 =
1
2

After simulations, output trace files, containing different 1+

(𝑚−𝑐)
𝑓·𝑐

vehicle trip information on waiting times, and running times,

TABLE II. SQV Statistic Guide
is generated. From the different trip information collected,
the performance indicators are collected for benchmarking. SQV Value Indicator
Calculations for the different algorithm performance SQV ≥ 0.90 Very good match
indicators were adapted from Magueflor et. al [10]. 0.85 ≤ SQV < 0.90 Good match
0.80 ≤ SQV < 0.85 Acceptable match
a)​ Queue Length - Total number of vehicles queued up in an
SQV < 0.8 Unacceptable match
incoming segment of an intersection (Unit: veh)
b)​ Queue Time - Total time spent by each vehicle in a queue
throughout a trip (Unit: s) 2)​ Benchmarking
c)​ Vehicular Flow Rate - Total number of vehicles that pass To assess the performance of the implemented control
through an intersection divided by time (Unit: veh/hr) algorithms, benchmarking of the traffic signal control (TSC)
d)​ Number of Cars Serviced - Number of cars arriving in
their destination over a time frame (Unit: veh) schemes was done against the current fixed-time scheme for
the traffic network. The fixed-time TSC was simulated over
3)​ Python-SUMO Integration the 4-hour period, and the performance indicators were
The traffic simulation in SUMO and the MPC algorithm compared against the adaptive TSCs, in both centralized and
in Python were integrated via SUMO’s Traffic Control decentralized approaches. The main sections of the
Interface(TraCI). TraCI, operating over a TCP connection, performance benchmarking were focused on traffic peak
allows the Python-based MPC client to connect to the hours (6AM-10AM). Additionally, the performance of the
SUMO server socket. This connection was initiated by control schemes were also tested for off peak hours
setting the SUMO HOME environment variable, importing (11AM-3PM).
TraCI into the Python script, and invoking the start function
to connect to sumo-gui with the fixed time configuration file IV.​ RESULTS AND DISCUSSIONS

a)​ Store-and-Forward Model This section presents the results of the traffic
To load cars into each node within the simulation, TraCI simulations in SUMO. Validation of simulation results
counted all cars halted on the roads around the intersection. against the MMDA-collected data is presented through the
Cars are considered halted when their speed is below a SQV statistic. The performance indicators of each traffic
threshold of 0.1m/s. For roads whose vehicles may take scheme configuration are compared, for both peak and
different routes from, a helper function randomly assigned off-peak demand profiles as well as their respective
vehicles to each of the possible routes, weighted to the improvement metrics. Plots for the performance indicators
previous patterns of traffic. over 15-minute fixed windows are also presented to
visualize the trends over the simulation time. As well as
b)​ Cell Transmission Model green times of each phase in relation with the demand
To generate cells inside the simulation, Lane Area (E2) composition of each road. A synthesis is provided in the
detectors were used. These detectors were meticulously last section, summarizing the relevant results and insights
placed over the different lanes depending on the cell length. gained from the traffic simulations.
Using TraCI, different cell parameters were collected from
the simulation such as the number of cars inside the cell, as A.​ Verification of Simulation-generated Demand vs
well as cell occupancy. These parameters were essential in MMDA Demand Data
computing in the decision function for the control algorithm. All the traffic simulations have undergone an SQV
statistical analysis for the traffic demand generated by
TraCI is also used to actuate the traffic signal based on SUMO for each hour of the 4-hour duration of the
triggers inside the MPC python code for the cell simulation. Since there are two proposed traffic control
transmission model. Decision functions are used to execute schemes with each having a centralized and decentralized
phase adjustments and cut down green times of the current version in addition to the fixed time TSC, there were a
active phase once delays are seen to be more severe in other total of 5 simulations to validate. Actual demand vs
phases. simulated demand for all TSCs for peak hours are shown
D.​ Validation Phase in Table III while Tables IV shows the result of its SQV
validation. Tables V and VI show the results of the SQV
1)​ Traffic Model Validation validation for off peak hours.
Model validation was accomplished by doing SQV
TABLE III. Actual vs Simulated Demand at Peak Hours
statistics between the simulated demand and actual demand
to ensure that the simulation configuration appropriately Demand (vehicles)
Time Simulated
matches that of the actual traffic conditions of both the
Actual Store-and-Forward Model Cell Transmission Model
intersections. SQV normalizes values between 0 and 1 and Start End
Fixed
Centralized Decentralized Centralized Decentralized
introduces a scaling factor f, allowing it to be adapted to 6:00 7:00 10357 10358 10358 10358 10358 10061
3
various situations. In this project, f was set to 10 to 7:00 8:00 11737 11664 11736 11687 11630 10312
compare the simulator’s per-hour traffic volume with the 8:00 9:00 12574 12360 12330 12421 12339 12883
procured hourly demand data. The formulation of SQV is 9:00 10:00 9638 9602 9645 9646 9905 10490
shown in Equation (5), and Table II provides interpretations TOTAL 44306 43984 44069 44112 44232 43746
for different SQV values. Backlog 322 237 194 74 560

©2024 IEEE
 the Fixed-time scheme as shown in Table VIII and X. For
TABLE IV. Validation of All TSC Simulation at Peak Hours
further visualization of the tables refer to Appendix A and
SQV B at the end of the paper.
Time
Store-and-Forward Model Cell Transmission Model
Fixed TABLE VII. Performance Comparison for all TSCs for Peak Hours
Start End Centralized Decentralized Centralized Decentralized
Performance Store-and-Forward Cell Transmission
06:00 07:00 0.9997 0.9997 0.9997 0.9997 0.9158 Fixed
07:00 08:00 0.9791 0.9997 0.9856 0.9697 0.7062 Indicator Centralized Decentralized Centralized Decentralized
Average queue length (veh) 12.37 11.85 16.06 11.49 20.50
08:00 09:00 0.9431 0.9356 0.9586 0.9378 0.9198
Average time in queue (s) 96.06 93.34 125.92 89.69 161.86
09:00 10:00 0.9885 0.9978 0.9974 0.9208 0.7847
Average vehicular flow rate (veh/hr) 19365.75 19391.25 19287.50 19368.50 19277.00
4-hour duration 0.9849 0.9889 0.9909 0.9965 0.9741
Vehicles serviced 43493 43636 43490 43712 43332
For all five TSCs during peak hours, the SQV is
TABLE VIII. Improvement against Fixed-Time TSC for Peak Hours
consistently high, falling under the very good match
classification in the SQV value indicator. The SQV validates Improvement
Performance
all the proceeding results of the proposed traffic control Store-and-Forward Cell Transmission
Indicator
Centralized Decentralized Centralized Decentralized
strategies since it indicates that the simulations are effective
Average queue length (veh) 4.26% -29.77% 7.16% -65.67%
in mirroring the traffic situation at the selected intersections,
Average time in queue (s) 2.83% -31.08% 6.64% -68.49%
according to the data provided by the MMDA. Since Average vehicular flow rate (veh/hr) 0.13% -0.40% 0.01% -0.46%
vehicular volumes are generated on an hourly basis, the
Vehicles serviced 143 -3 219 -161
simulated vehicles are capped by the actual value, and the
simulated values exceed the actual values in cases where TABLE IX. Performance Comparison for all TSC for Off-Peak Hours
vehicle backlogs from previous hours are completed. A Performance
Fixed
Store-and-Forward Cell Transmission

limitation of the simulator is that it distributes the vehicular Indicator Centralized Decentralized Centralized Decentralized

volume over an hour equally, which fails to capture the Average queue length (veh) 6.35 6.18 8.74 5.38 16.69
randomness in actual traffic. Average time in queue (s) 61.21 59.30 84.28 51.79 166.91
TABLE V. Actual vs Simulated Demand at Off-Peak Hours Average vehicular flow rate (veh/hr) 15927.5 15916.5 15774.5 15884.75 147580
Vehicles serviced 35885 35867 35674 35955 33449
Demand (vehicles)
Time Simulated TABLE X. Improvement against Fixed-Time TSC for Off-Peak Hours
Actual Store-and-Forward Model Cell Transmission Model Improvement
Fixed Performance
Start End Centralized Decentralized Centralized Decentralized Store-and-Forward Cell Transmission
Indicator
11:00 12:00 10357 10358 10358 10358 10358 10061 Centralized Decentralized Centralized Decentralized
12:00 13:00 11737 11664 11736 11687 11630 10312 Average queue length (veh) 2.54% -37.72% 15.20% -163.04%
13:00 14:00 12574 12360 12330 12421 12339 12883 Average time in queue (s) 3.12% -37.69% 15.39% -172.68%
14:00 15:00 9638 9602 9645 9646 9905 10490 Average vehicular flow rate (veh/hr) -0.07% -0.96% -0.27% -7.34%
TOTAL 44306 43984 44069 44112 44232 43746 Vehicles serviced -18 -211 70 -2436
Backlog 11 11 11 11 1853 The fixed time performance indicators are obtained from
TABLE VI. Validation of All TSC Simulation at Off-Peak Hours the simulation, served as the baseline metrics to be improved
SQV on and were compared to the parameters of the algorithms.
Time
Fixed
Store-and-Forward Model Cell Transmission Model Due to the fact that the Bayesian optimization algorithm
Start End Centralized Decentralized Centralized Decentralized
of the centralized S&F model prioritized maximizing flow
11:00 12:00 0.9996 0.9996 0.9996 0.9996 0.9996 rate, the flow rate is far and away the best performing
12:00 13:00 0.9996 0.9996 0.9996 0.9996 0.9996 metric. Furthermore, it out-performed the decentralized
13:00 14:00 0.9997 0.9997 0.9997 0.9997 0.7910 system due to the fact that the centralized system ensured
14:00 15:00 0.9956 0.9956 0.9956 0.9956 0.7569 that phase 1 occurs simultaneously throughout both
4-hour duration 0.9994 0.9994 0.9994 0.9994 0.9113 intersections, leading to an increase in efficiency.
The same is applicable for the off peak simulations as
most time slots fall under the very good match category. A The CTM Centralized model boasts an improvement in
few time slots produced unacceptable matches due to its queue time and queue length. For its flow rate however,
queueing problems in the control. Some time slots in the the result is almost similar to the fixed time. This behavior is
off-peak setting also produce closer matches due to the expected as CTM tries to cater phases that have more delay
lessened occurrence of backlogs because of the off-peak time, and exhaust the cars in that phase. Overall, the CTM
schedule. In summary, all TSC simulations are deemed Centralized serviced more vehicles than the fixed time
acceptable, and the number of generated vehicles aligns with model. The CTM Decentralized TSC, compared to the
the actual demand data across all simulations. Centralized model, suffered in all performance indicators.
This was caused by the asynchronous control of the two
B.​ Performance Indicators Comparison intersections, causing vehicles to queue in the middle of
In this section, a comparative analysis of the Katipunan road. In summary, the Store and Forward model
performance indicators between the existing fixed-time produced the highest flow rate, while the Cell Transmission
traffic control system and the model predictive control Model produced the greatest improvement in terms of queue
(MPC)-based traffic control schemes are performed. Table length, queue time, and vehicles serviced. Both centralized
VII and IX. highlights best performances among the 5 TSC configurations also serviced more cars during the peak
schemes for both peak and off-peak hours, respectively. period. For off-peak hours, the CTM implementation tends
The performance metrics comparison of Table VII and IX to suffer more as the lower number of volume makes the
are used to calculate the improvement of the TSCs against control switch at a faster rate. This decentralized switching

©2024 IEEE
creates a larger gap between the synchronicity of the two
intersections, making the queues longer and the evaluation
more unstable.
C.​ Time Series Analysis
This section provides an analysis of the traffic
simulation throughout the 4-hour simulation. The proposed
TSCs model parameters are first compared to the current
fixed-time TSC by plotting the values of the performance
indicators over time. Afterwards, to visualize the proposed Fig. 5. 15-Minute Window Queue Time Comparison for Peak Hours
TSC schemes performance, an inspection on how the
algorithm chooses green times for certain traffic demand
composition is conducted.
1)​ Comparison of MPC-based TSCs vs. Current
Fixed-time TSC
Fig. 3 below shows the average queue length in vehicles
across all TSC schemes, following an upward trend for the
first three hours that follows the increasing number of
vehicles from the actual demand data.The graphs show Fig. 6. 15-Minute Window Queue Time Comparison for Off-Peak Hours
prominent peaks in the 7:00 AM to 9:00 AM interval as this In terms of flow rates, all TSCs experienced a sudden
is the time where the demand is the highest. Store and increase from 8:00 to 9:00. This is due to the fact that the
Forward allowed for longer queues as no fixed cycle times largest portion of vehicles spawned were serviced during
were applied to the four phases. those times. In this area, no MPC control managed to
Fig. 4 shows the off-peak queue lengths for all TSC produce results better than the fixed time. Fig. 7 and Fig. 8
schemes. Compared to peak times, the decentralized models illustrate the flow rate behavior. Yet, observing the flow rate
tend to have higher averages due to the asynchronous nature, graph of both the CTM models, it is noticeable that the flow
with a higher average than the Fixed TSC, where CTM has rate does not swing as compared to both the Fixed time TSC
the highest. and SF TSC. This is caused by CTM’s behavior to converge
the green time it allows the network based on the ratio of a
considerably full road. The progression of the flow rate is
more subtle as the prediction region still counts the number
of cars that cannot outflow in the next time step.

Fig. 3. 15-Minute Window Queue Length Comparison for Peak Hours

Fig. 7. 15-Minute Window Flow Rate Comparison for Peak Hours

Fig. 4. 15-Minute Window Queue Length Comparison for Off-Peak Hours

The trend from queue lengths also follows for average
queue times across the different TSC schemes as seen in
Fig. 5 And Fig. 6, the centralized approach for CTM
produced the shortest queue lengths and queue times as its Fig. 8. 15-Minute Window Flow Rate Comparison for Off-Peak Hours
control approach aimed to minimize delays experienced 2)​ Green Times vs. Traffic Volume Distribution
across phases. The Store and Forward-based TSCs tend to To gain insight into the substantial enhancements in
have higher queue times compared to CTM due to the average queue length and queue time achieved by the
prioritization of flow rate, leading to larger queues in MPC-based TSC, an examination of the green times applied
intersections. Regardless, it still performed better than the at the intersection was conducted. In the subsequent series of
Fixed time. plots, the green times and the road demand composition
during both peak and off-peak hours over a 4-hour duration
are shown. This allowed for an appreciation of how the
MPC dynamically adjusted timer settings to accommodate
real-time traffic demands.

©2024 IEEE
 Fig. 9 shows the demand composition during the peak
hours of traffic. Due to the number of vehicles coming from
upper and lower Katipunan, longer green times are assigned
to the phases with those roads involved (typically Phase 1
and 4).

Fig. 12. Green Times Applied SF Decentralized TSC at KUF intersection

Fig. 13 shows the green times applied in the CTM

Centralized configuration. The green times for phase 1
follow the increasing trend of the traffic volume from
Fig. 9. Demand Composition at 6:00 AM to 10:00 AM 6-9AM. As assignments of green time are based on delay
times experienced by queues, phases change much more
Fig. 10 shows the green times applied across the four
quickly. And while this may be effective in minimizing
phases under the Store and Forward Centralized TSC.
queue lengths and queue times, it has implications on
Minimum and maximum green times are set per phase to
driving experience and is reliant on fast reaction by drivers
avoid queue starvation. As the optimal green times are
in an intersection.
computationally expensive, a minimum car checker is
utilized for Store and Forward, which applies a minimum
green time of 25 seconds whenever the number of cars
queued up are less than a set threshold. This explains some
of the phases with small green times.

Fig. 13. Green Times Applied for CTM Centralized TSC

The decentralized green times of the CTM TSC show
different behavior for the two intersections. As the control
Fig. 10. Green Times Applied for 4 hours in SF Centralized TSC now relies on traffic volume on individual road connections
of an intersection, the KBT intersection prioritized the
Fig. 11 and Fig. 12 shows the green times applied across exhaustion of cars along the B. Gonzales Road, while the
the four phases under the Store and Forward Decentralized KUF intersection still gives the highest green times to the
TSC. With green time computations now independent per Katipunan phases. This asynchronous behavior greatly
intersection, larger green times are applied for phase 1 of the increases the queue in the middle katipunan road, causing
KUF intersection as the cars from this road account for 39% the KUF intersection to hoard the green time for the Phase 1.
of the vehicular volume. The max green time for Phase 4 in
KBT is also assigned more frequently as it is also a major
source road in the intersection. This explains some of the
phases with small green times. Due to the “locality” of the
decentralized approach, it is able to assign minimum green
times more frequently to different phases as it does not need
to consider the cars queued up in the other intersection.
Phase 3 in KBT receives minimum green times more
Fig. 14. Green Times Applied CTM Decentralized TSC at KBT intersection
frequently because some of the cars queued up for that phase
are serviced during its Phase 4, while Phase 3 in KUF has
more dynamic green time assignments due as it has to cater
to more volume of cars.

Fig.15. Green Times Applied CTM Decentralized TSC at KUF intersection

D.​ Vehicular Queuing Behavior
Although the expected queueing behavior of the
Fig. 11. Green Times Applied SF Decentralized TSC at KBT intersection motorcycles were replicated through the application of
sublanes, the trade-off for this specification was the
occasional emergence of two cars queuing in one lane.
While this is not the usual behavior for traffic, instances
like this can represent the switching of lanes by a car as it
enters a queue. The left of Fig. 16 depicts the expected

©2024 IEEE
real-life behavior of motorcycles, while the right illustrates _Concept_Jeepney_NAMA.pdf
[3]​ MMUTIS STUDY TEAM, “Metro Manila Urban
the actual behavior of motorcycles in the simulation and Transportation Integration Study Technical Report No. 8,” Mar.
Fig. 18 shows two cars queuing in a single lane in the 1999. Available:
simulation. https://openjicareport.jica.go.jp/pdf/11580545.pdf
[4]​ K. Vergel and T. Yai, “ANALYSIS OF ROAD TRAFFIC
FLOW AND TRAFFIC ENVIRONMENT IN METRO
MANILA,” Aug. 2018. Available:
https://ncts.upd.edu.ph/tssp/wp-content/uploads/2018/08/Vergel
00.pdf
[5]​ S. Lee, S. C. Wong, and P. Varaiya, “Group-based hierarchical
adaptive traffic-signal control Part II: Implementation,”
Transportation Research Part B: Methodological, vol. 104, pp.
Fig. 16. Multiple Motorcycles Queued Side-by-side in Aurora Blvd. 376–397, Oct. 2017, doi:
East.(left) Taken from Google Maps[10] and Actual Queuing Behavior of https://doi.org/10.1016/j.trb.2017.08.009.
Motorcycles(right) [6]​ L. Ma, B. Xue, and J. Wu, “Centralized Traffic Signal Control
for Multiple Intersections based on Sequence-to-Sequence
model and Attention Mechanism,” 2021 IEEE International
Intelligent Transportation Systems Conference (ITSC), Sep.
2021, doi: https://doi.org/10.1109/itsc48978.2021.9564818.
[7]​ Tamás Tettamanti, I. Varga, Balázs Kulcsár, and J. Bokor,
“Model predictive control in urban traffic network
management,” 2008 16th Mediterranean Conference on Control
and Automation, Jun. 2008, doi:
Fig. 17. Unusual Queuing Behavior of Cars in SUMO https://doi.org/10.1109/med.2008.4602084.
V.​ CONCLUSION [8]​ F. Abbracciavento, F. Zinnari, S. Formentin, A. G. Bianchessi,
and S. M. Savaresi, “Multi-intersection traffic signal control: A
In conclusion, the Centralized Store and Forward decentralized MPC-based approach,” IFAC Journal of Systems
and Control, vol. 23, p. 100214, Mar. 2023, doi:
Model managed to produce a higher vehicular flow rate,
https://doi.org/10.1016/j.ifacsc.2022.100214.
and lower average queue lengths and queue times than the [9]​ H. K. Lo, “A novel traffic signal control formulation,”
fixed time approach. The Centralized Cell Transmission Transportation Research Part A: Policy and Practice, vol. 33,
Model managed to achieve lower average queue lengths no. 6, pp. 433–448, Aug. 1999, doi:
https://doi.org/10.1016/s0965-8564(98)00049-4. ‌
and queue times than the fixed time approach. The
[10]​ E. S. T. Go, F. M. M. Magueflor, and J. L. R. Uy, "Smart
Decentralized Cell Transmission Model managed to Traffic - Application of Data-Driven Control in Local
achieve lower average queue lengths and queue times than Timer-Based Stoplights," unpublished undergraduate thesis,
the fixed time approach. Most simulations were deemed at Electrical and Electronics Engineering Institute, University of
the Philippines Diliman, Quezon City, Metro Manila,
least acceptable matches to real data if not better via the
Philippines, 2023
SQV statistic. In order to further improve this study in the
future, conducting whole-day simulations would give a
more comprehensive overview of the performance of each
model. Introducing fixed cycle times for the Store and
Forward model would allow it to avoid sacrificing queue
times for higher flow rates. Optimizing other parameters
for SF, and optimizing the decentralized CTM is also
recommended for future works. Additionally, expanding
the study to include more intersections will provide a
comprehensive analysis of the traffic control strategies.
Acknowledgments
To our adviser, Dr. Jhoanna Pedrasa, for guiding us
throughout the research. Her patience and dedication
through our consultations made it possible for us to
complete our project. To Eisler Spencer Go, Francis
Magueflor, and Lendl Uy, for paving the way and for
providing guidance in this endeavor. To the Department of
Science and Technology - Science Education Institute, for
granting us support to accomplish this project. To our
friends and families, for giving support and encouragement
throughout the tough times. And to the Filipino citizens who
spend hours and hours in traffic, we dedicate this project to
you.
References
[1]​ TomTom International BV20, “Manila traffic report | TomTom
Traffic Index,” Available: www.tomtom.com, 2023.
https://www.tomtom.com/traffic-index/manila-traffic/
[2]​ Department of Transportation (DOTr) of the Philippines,
“Transforming Public Transport in the Philippines” Available:
https://changing-transport.org/wp-content/uploads/Full_NAMA

©2024 IEEE
APPENDIX A
Visualization for the performance indicators comparison as
well as the improvement for peak duration.

Fig. 18. Bar graph for Average Queue Length Comparison (Peak) Fig. 21. Bar graph for Average Queue Length Improvement (Peak)

Fig. 19. Bar graph for Average Queue Time Comparison (Peak) Fig. 22. Bar graph for Average Queue Time Improvement (Peak)

Fig. 20. Bar graph for Average Vehicular Flow Rate Comparison (Peak) Fig. 23. Bar graph for Average Vehicular Flow Rate Improvement (Peak)

©2024 IEEE
APPENDIX B
Visualization for the performance indicators comparison as
well as the improvement for off-peak duration.

Fig. 24. Bar graph for Average Queue Length Comparison (Off-Peak) Fig. 27. Bar graph for Average Queue Length Improvement (Off-Peak)

Fig. 25. Bar graph for Average Queue Time Comparison (Off-Peak) Fig. 28. Bar graph for Average Queue Time Improvement (Off-Peak)

Fig. 26. Bar graph for Average Vehicular Flow Rate Comparison Fig. 29. Bar graph for Average Vehicular Flow Rate Improvement
(Off-Peak) (Off-Peak)

©2024 IEEE