GREE-COCO: Green Artificial Intelligence Powered Cost Pricing

Models for Congestion Control

Meghana Kshirsagar

, Tanishq More

, Rutuja Lahoti

, Shreya Adgaonkar

, Shruti Jain

Conor Ryan

and Vivek Kshirsagar

Biocomputing Developmental Systems, University of Limerick, Ireland

Department of Information Technology, Government College of Engineering, Aurangabad, India

Department of Computer Science, Government College of Engineering, Aurangabad, India

Keywords: Intelligent Transportation, Traffic Management, Congestion Control, Cost Pricing, YOLOv3, F-RCNN,

Mask R-CNN, Energy Efficient.

Abstract: The objective of the proposed research is to design a system called Green Artificial Intelligence Powered

Cost Pricing Models for Congestion Control (GREE-COCO) for road vehicles that address the issue of

congestion control through the concept of cost pricing. The motivation is to facilitate smooth traffic flow

among densely congested roads by incorporating static and dynamic cost pricing models. The other

objective behind the study is to reduce pollution and fuel consumption and encourage people towards

positive usage of the public transport system (e.g., bus, train, metro, and tram). The system will be

implemented by charging the vehicles driven on a particular congested road during a specific time. The

pricing will differ according to the location, type of vehicle, and vehicle count. The cost pricing model

incorporates an incentive approach for rewarding the usage of electric/non-fuel vehicles. The system will be

tested with analytics gathered from cameras installed for testing purposes in some of the Indian and Irish

cities. One of the challenges that will be addressed is to develop sustainable and energy-efficient Artificial

Intelligence (AI) models that use less power consumption which results in low carbon emission. The GREE-

COCO model consists of three modules: vehicle detection and classification, license plate recognition, and

cost pricing model. The AI models for vehicle detection and classification are implemented with You Only

Look Once (YOLO) v3, Faster-Region based Convolutional Neural Network (F-RCNN), and Mask-Region

based Convolutional Neural Network (Mask RCNN). The selection of the best model depends upon their

performance concerning accuracy and energy efficiency. The dynamic cost pricing model is tested with both

the Support Vector Machine (SVM) classifier and the Generalised Linear Regression Model (GLM). The

experiments are carried out on a custom-made video dataset of 103 videos of different time duration. The

initial results obtained from the experimental study indicate that YOLOv3 is best suited for the system as it

has the highest accuracy and is more energy-efficient.

https://orcid.org/0000-0002-8182-2465

1 INTRODUCTION

Universally there is a severe issue of growing traffic

congestion causing severe traffic problems which is

contributing to the rise of pollution in cities and

towns, leading to emissions of carbon monoxide gas

and smog. Further, the emissions from fossil fuel

vehicles produce greenhouse gases contributing to

warmer temperatures. Hence a need for traffic

analysis is essential for exploring the possibility of

building large-scale infrastructures needed for public

transportation. Such systems, which are critical

components of any government, can be utilised to

provide communities with the most effective,

functional, and environmentally friendly

transportation. This research work’s primary

motivation is to design a system that collects traffic

data for analysing traffic volume measurements

from video recordings in different weather

conditions. To alleviate congestion problems,

various states and legislatures worldwide have

implemented cost pricing systems with the Pricing

strategy (U.S. Department of Transportation, 2008)

to be either Static Pricing (SP) or Dynamic Pricing

(DP). SP is fixed throughout whereas DP changes

according to various factors like timestamp and

traffic count. Singapore (DP) (Ye and S, 2012)

916

Kshirsagar, M., More, T., Lahoti, R., Adgaonkar, S., Jain, S., Ryan, C. and Kshirsagar, V.

GREE-COCO: Green Artiﬁcial Intelligence Powered Cost Pricing Models for Congestion Control.

DOI: 10.5220/0010261209160923

In Proceedings of the 13th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2021) - Volume 2, pages 916-923

ISBN: 978-989-758-484-8

implements Radio Frequency Identification (RFID)

systems, while cities like London (SP) and

Stockholm (Variable Pricing) (Börjesson and

Kristoffersson, 2018) implement Camera-based

congestion pricing systems (AECOM Consult team,

2006). Variable Pricing depends on changing the

charging rate only during the peak periods. In this

paper, the authors propose a combined approach that

applies Static Pricing for peak hours and Dynamic

Pricing for the non-peak hours. It also focuses on the

development of vehicle detection and classification

system that is energy efficient and hence contributes

towards the development of ‘Green AI’ (Schwartz et

al., 2019) sustainable models.

2 METHODS

The authors in (Song et al., 2019) proposed a

segmentation method consisting zones placed into

the YOLO v3 architecture to detect the vehicle type

and location that obtains the vehicle trajectories. The

authors in (Tourani et al., 2019) discuss a robust

method for detecting vehicles in video frames based

on Faster RCNN instead of the Residual neural

network (Resnet-50). In (Al-Ariny et al., 2020) used

the Mask RCNN instance segmentation model for

performing vehicle detection. In their proposed

approach, once objects are detected, corner points

are extracted and then tracked. In (Dorbe et al.,

2018), the authors built a system that performs

segmentation on an input image for detecting the

license plate and vehicle type classification. In

(Clements et al., 2020), the authors discuss in detail

the technologies that can be implemented for

congestion pricing applications, such as Vehicles-

miles-travelled, Zone-based tolling and Credit-based

pricing.

3 PROPOSED SYSTEM

ARCHITECTURE

Figure 1 describes the proposed integrated system

for congestion control. The video is captured and

given as an input to extract frames for performing

vehicle detection. The vehicle detection module is

implemented with three algorithms- YOLOv3 (Song

et al., 2019), Faster RCNN (Tourani et al., 2019) and

Mask RCNN (Al-Ariny et al., 2020). The detected

vehicle’s image is captured and further processed for

classification of the vehicle into five classes (car,

Figure 1: Proposed Architecture for GREE-COCO.

GREE-COCO: Green Artiﬁcial Intelligence Powered Cost Pricing Models for Congestion Control

917

bus, truck, bicycle, motorcycle), and license plate

detection and recognition. Based on the video

timestamp, the categorisation of peak and non-peak

hours is done. Along with the vehicle classification,

we calculate the total and individual count of the

vehicle to calculate the congestion charge for non-

peak hours. In contrast, the cost for peak hour is

static. The user’s details of the recognised license

plates of all vehicles are retrieved from the Regional

Transport Office (RTO) database, the Indian

government organisation responsible for maintaining

a database of drivers and vehicles for various states

of India, for which the Indian Government has

granted us access. Similarly, the detail of Irish

vehicles owners is accessed from the Transportation

Department. Based on the information from the

database, if a vehicle is classified as an emergency

service vehicle, it is exempted from the congestion

charge. If the vehicle is a taxi, then a separate

minimal cost is decided and sent via Email, and for

all other vehicle users, an Email is sent with details

of the congestion charge.

3.1 Vehicle Detection and Counting

In order to perform Vehicle Detection, Linear filters

are used for video pre-processing to correct the non-

uniform lights or colour and intensity adjustment.

Convolution operation performs linear filtering,

which is a technique in which the value of output

pixel is linear combination of the values of

neighbouring input pixel. The obtained weight

matrix is called as convolution kernel. The initial

step is to load an input video as a number of images.

The splitting frame rate can be monitored. To reduce

the noise, we need to blur the input image with

Gaussian Blur filter then convert it to grey scale.

Different morphological operations are applied to

enhance the edges of the image. To reveal the object,

we binarize the image then adaptive thresholding is

applied to eliminate irrelevant important

information. This is done to remove all the other

entities other entities except from image.

This module detects, recognises, and tracks the

vehicles in the video frames and classifies them into

one of the five classes as mentioned above. The

accurate selection of a region of interest (ROI) is

vital to decrease the false positives in the detection

and classification of vehicles. ROI provides the

flexibility of just working within a particular area

instead of manipulating the whole video. When a

specific vehicle is present in the ROI, then it gets

detected and the count increases. The Deep Sort’

algorithm (Wojke et al., 2017) is implemented for

calculating the count of vehicles by tracking vehicle

movements within a tracking zone labelled as a

virtual loop. This algorithm is known for the quality

real-time object tracking and can perform both

single object tracking and multiple object tracking.

For further analysis, the total count of vehicles is

stored in a database. The statistics of vehicles are

then given as an input to the dynamic congestion

pricing model to determine the congestion cost.

3.2 Cost Pricing Model

According to the cost pricing model, the user will be

charged for using a road as defined by its priority

level and time of usage. The factors used to calculate

the cost are the time of the day, the priority of the

road, season, and the vehicle class. A priority is

assigned to all types of roads, congested as well as

non-congested. The priority is directly proportional

to the number of vehicles on that road. The cost of

peak hours is constant, while that of the non-peak

hours is dynamic. The non-peak hour cost will be

calculated based on vehicle count at that particular

instance of time.

3.2.1 Peak Hours

For Peak Hours, the static prices will be charged to

all vehicles. Charges for trucks and buses will be

higher than that of cars and motorcycles. Emergency

vehicles such as ambulance, fire-trucks will be

exempted from this system. Moreover, Government

vehicles (for e.g. Police Department vehicles) will

not be charged.

3.2.2 Non-peak Hours

We have divided the total traffic count for a

particular hour into different classes. Each class has

a class width of 150. The first class starts with a

threshold count of 100 vehicles (e.g. 100 - 250 1st

vehicle count class). If the total count of vehicles at

that particular hour is less than threshold count, then

no vehicle will be charged.

A, B, C represents High, Medium, and Low

Priority roads, with weight value 30, 25 and 20

respectively. The values of A, B and C were

finalised after performing a series of experiments for

calculating the total congestion cost. The above road

priority values were finalized such that the total

congestion cost for the vehicles should lie between

(0, 30) Indian Rupee (INR).

Consider Total Traffic count at a given

timestamp as T, and Total Individual Count of Car,

ICAART 2021 - 13th International Conference on Agents and Artiﬁcial Intelligence

918

Bus, Truck, and Motorcycle are C

, C

and C

respectively.

We Define,

𝑅𝑖 =

𝐶𝑖

𝑇

(1)

Where R

= Ratio of particular vehicle class

contributing to traffic, where i∈ (1,4)

..R

representing the ratio of car, bus, truck, and

motorcycle contributing to traffic, respectively.

Ci = Count of particular vehicle class, here i∈ (1,4)

We have,

∑

𝑅𝑖





(2)

We calculate the cost factor for a vehicle as

𝐶𝑓 = (𝑀𝑖 × 𝑅𝑝 )

(3)

Where C

= Cost Factor;

= Mean of the class in which Total traffic lies;

and R

= Road Priority value.

So,

𝑀𝑖 =

𝑈𝑖 𝑥 𝐿𝑖

(4)

Here U

= Upper frequency of the class in Total

traffic (T) lies while

= Lower frequency of the class in Total traffic (T)

lies.

𝐴

𝑖 =

𝑅𝑖 𝑥 𝐶𝑓

𝑇

(5)

∀ A

= Congestion cost of each vehicle class, where

i∈ (1,4) and

..A

represents the congestion cost of the car, bus,

truck, and motorcycle, respectively.

For electric vehicles, there is an incentive of a 0.7

reduction in the cost, as shown in equation 6.

𝐴

𝐸𝑖 = (

𝐴

−(

𝐴

𝑖×0.7))

(6)

∀AE

= Congestion cost of each electric vehicle

class, where i ∈ (1,4) and AE

..AE

represents

electric vehicle congestion cost of the car, bus, truck

and motorcycle respectively.

The cost for a particular road, for example, road

‘A,’ is calculated using equations 1 to 6. The vehicle

detection and classification module are used to

detect vehicles on road ‘A’. After classifying and

finalising the cost of the vehicles, the image is

captured and given as an input to the next module,

License Plate Detection and Character Recognition.

3.3 License Plate Detection and

Recognition

Algorithm 1: DLP: Detection of License Plate.

Input: Coloured Image A of a vehicle

Output: License plate number

1: START:

2: image Input Image

3: grayscale convert image to gray

4: for each pixel in image:

5: pixel=pixel/255.0

6: end for

7: R_image resize (224,224)

8: Input R_image to Wpod network

9: License_Plates[]  save license plates

10: for license_plate in License_Plates:

11: x,y,x+h,y+w get coordinates of

license_plate

12: A  crop image with the coordinates x,y,x+h,y+w

13: G_A  cv2.cvtColor(A,

cv2.COLOR_BGR2GRAY)

14: M_A  Perform noise reduction and morphological

operations on G_A

15: Draw contours  cv2.findContours(binary,

cv2.RETR_EXTERNAL,

cv2.CHAIN_APPROX_SIMPLE)

16: for contours in Contours:

17: B  crop & save contours

18: Input cropped_contour to Letter &Digit recognition

model

19: A[]  append the predicted the letter

20: end for

21: end for

22: Display A

23: END

The module uses a set of a convolutional neural

network, computer vision techniques with OpenCV,

and character recognition as shown in Algorithm

DLP. The license plate character recognition is

performed using OpenCV methods such as changing

colour spaces, image filtering, edge detection and

image contours. The process is divided in three

steps. The first part implements a pre-trained model

Wpod Net (Silva and Jung, 2018), to detect and

extract the license plates from the vehicle images.

Wpod Net is known for its ability to detect multiple

license plates from a single frame. In the second

part we perform plate character segmentation.

Finally, the CNN recognises the digits from the

extracted license plate. The extracted license plate

image follows a series of transformations, as

illustrated in Figure 2.

GREE-COCO: Green Artiﬁcial Intelligence Powered Cost Pricing Models for Congestion Control

919

Figure 2: License Plate and Character Recognition in

different scenarios. (a) Vehicle, (b) Detected Plate, (c)

Gray LP, (d) Blur LP, (e) Binary LP, (f) Dilation LP,

(g)Character detection, (h) Digit Recognition.

The license plate number of the vehicle obtained

from the License Plate Recognition module is stored

in a database. License Plate database is connected to

the RTO database to fetch the user’s details. The

RTO database attributes are vehicle type,

registration number, date of issue, fuel or electric

based engine, and the user’s detail. The charges for

the user are sent on their Email.

4 RESULTS

This section will discuss the performance of all three

modules based upon accuracy and energy efficiency.

4.1 Vehicle Detecting and

Classification

Here we discuss the performance of the AI models.

4.1.1 Video Dataset

Table 1: Test Video Dataset.

Dataset

Time

Duration

No of

videos

No of

frames

D1 10 sec 30 7780

D2 15 sec 25 9780

D3 30 sec 17 13380

D4 1 min 8 12120

D5 2 min 12 37200

D6 5 min 11 87000

Total 103 167260

The vehicle detection and classification module were

tested on 103 videos, which were a combination

Figure 3: Vehicle Detection with (a) YOLOv3 (b) Mask RCNN (c) Faster RCNN.

Figure 4: Confusion Matrix for vehicle counting and classification with (a) YOLOv3, (b) Faster RCNN, (c) Mask RCNN.

ICAART 2021 - 13th International Conference on Agents and Artiﬁcial Intelligence

920

of public datasets (Chan and Vasconcelos, 2005)

(Wang et al., 2019) and YouTube videos of varying

time duration. In Table 1, each Dataset D1, D2 and

D3 consists of a total of 30, 25 and 17 videos of

10sec, 20sec and 30sec time duration respectively.

The total number of frames in the three datasets are

7780, 9780 and 13380. In a similar way, there are 8,

12 and 11 videos of 1 min, 2 min and 5 min in

dataset D4, D5 and D6.

4.1.2 Classification Accuracy

Figure 3(a), (b), (c) displays the output of vehicle

detection and classification modules with all the

three algorithms which are compared to identify the

best suited model. Figure 4 shows the confusion

matrices for vehicle counting and classification

models for all the three approaches which highlight

that Mask RCNN has the highest accuracy. Each cell

in the matrix describes the accuracy for a specific

class. For testing the model, the video datasets as

illustrated in Table 1, were used.

4.2 Cost Pricing with SVM and GML

The training dataset of the cost pricing model was

created using the NumPy library for 10,000

(465days) hours of data. The dataset consists of a

count of individual vehicle classes, the total count of

vehicle, road priority, and the expected particular

cost of each vehicle class. In the dataset the car

count lies between n (0, 400) per hour and, the bus

count is in the range (0, 80) per hour. The

Motorcycles, count varies from 0 to 550 while the

truck count lies in the range 0 to 4.

Figure 8 represents the data used for the training

SVM and GLM regression models, which predicts the

congestion charge for each vehicle class depending

upon peak and non-peak hours. The testing dataset

consisted of 1000 (41 days) hours of data.

4.3 Accuracy of Regression Models

The SVM and GLM models are trained on cost

pricing dataset as mentioned in (4.2) to predict the

congestion charge of vehicles at a particular hour.

Figure 5 (a) and 5 (b) illustrate the actual vs.

predicted congestion charge of Car, Motorcycle, Bus

and Truck for 100 hours respectively

4.4 License Plate Recognition

For recognition of a License Plate, we used a

publicly available dataset (Esther, 2018) on Kaggle.

The Data Set comprised 42,000 images. The dataset

Figure 5 (a): Expected vs. Predicted cost for different vehicle classes with SVM.

Figure 5 (b): Expected vs. Predicted cost for different vehicle classes with GLM.

GREE-COCO: Green Artiﬁcial Intelligence Powered Cost Pricing Models for Congestion Control

921

was split into the ratio of 80:20 for training and

testing the model. Mobile net pre-trained network

(Howard et al., 2017) was used to save

computational energy and time. The model was

trained for 50 epochs with a batch size of 64. The

training loss obtained was 0.0809, and the accuracy

was 0.9749.

4.5 Estimation of Energy Efficiency

The power consumed by the models on both

software and hardware platform was calculated,

which helped towards ranking the models according

to their energy consumption and carbon emission.

The hardware testing was performed on both CPU

and GPU platforms (Strubell et al., 2019). System

requirements for both CPU and GPU are X86_64

OS (Ubuntu 18.04) and 8GB RAM. Hardware

requirements for CPU are Packages, Core or

Noncore system and Dram.

The hardware used for the GPU was an “Nvidia-

GTX, 1050” with the latest drivers from Nvidia June

2020 release. The Python library ‘Energy Usage’

was used to calculate the energy consumption of

models which is based on ‘Running Average Power

Limit’ (RAPL) technology and supports the Nvidia-

smi program. The testing was performed on various

videos of duration 10sec, 20 sec, 30sec, and 40sec

on the YOLOv3 model, Faster RCNN, and Mask

RCNN model. The authors calculated the energy

consumed by the model based on baseline wattage

that calculates the computer's average power

required to initiate a process. Total wattage

calculates the energy consumed by the computer's

average power usage while the process runs. Process

wattage is the difference between the baseline and

total, highlighting the usage solely from the specific

process evaluated, and the process used is the total

amount of energy required to run the model. Carbon

emission is calculated by converting the kWh to

Carbon Dioxide (CO

) based on the energy mix of

the location as the emission differ based on the

country’s energy mix (García-Martín et al., 2019).

Figure 6 (a) and 6 (b) depict the total energy

consumed by different operations based on CPU and

GPU, respectively. Here the baseline wattage of all

the three models lies between 3.00 and 5.80 watts.

We can infer that more the time required to run total

processes in a model is directly proportional to CO

emitted by an individual model. The main difference

between the CPU and GPU performance is that GPU

requires less computational power to run a model

than CPU.

The energy estimation results are preliminary

and based on short samples of 103 videos as of now.

In future we will explore this avenue in more detail

by investigating the possibilities of parallelising the

process. This will be important especially when we

will test real time traffic data where millions of

videos may be needed to be processed and the idea

of incorporating distributed computing may lead

towards more efficient speeds while maintaining the

accuracy.

Figure 6 (a): The total amount of energy required to run a

model on the CPU.

Figure 6 (b): The total amount of energy required to run a

model on GPU.

5 CONCLUSIONS AND FUTURE

SCOPE

The research work proposes the design of an

intelligent transportation system named GREE-

COCO for managing road vehicle congestion. The

working prototype of the system is designed by

incorporating the approach of cost pricing. Traffic

data analytics is performed on 103 videos used in

experimentation. The authors tested the working

prototype on three popular multi-object detections

AI models like YOLOv3, F-RCNN and Mask-

RCNN. The comparative analysis was performed in

terms of accuracy and energy efficiency among

ICAART 2021 - 13th International Conference on Agents and Artiﬁcial Intelligence

922

models. The dynamic cost pricing model was tested

using regression models, SVM and GLM. It was

shown that YOLOv3 is best among three in terms of

accuracy, speedup while also being energy efficient.

Future work of the authors includes real-time testing

for camera-based approach in several Indian and

Irish cities. To overcome implementation challenges

like Public Acceptance, initially, the vehicles will be

charged with minimal cost. The problems related

with improper License Plate (Muddy, Snowy, Blank

or Damaged License Plate), will be mitigated by the

implementation of the RFID system in authors future

work. Moreover, Social and Political barriers can

also be overcome by making the system robust and

sustainable.

ACKNOWLEDGEMENTS

This work is supported under World Bank Technical

Education Quality Improvement Programme Seed

Grant (Government. College of Engineering,

Aurangabad) and partially supported by Science

Foundation Ireland grant #16/IA/4605.

REFERENCES

US Department of Transportation. 2008. Technologies

That Enable Congestion Pricing A Primer.

https://ops.fhwa.dot.gov/publications/fhwahop08042/f

hwahop08042.pdf

Ye, S. 2012. Research on Urban Road Traffic Congestion

Charging Based on Sustainable Development. In

Physics Procedia, 24, 1567-1572.

Börjesson, M., Kristoffersson, I. 2018. The Swedish

congestion charges: Ten years on. Transportation

Research Part A: Policy and Pra ctice, 107, 35–51.

AECOM Consult team. 2006. International Urban Road

Pricing,

http://bic.asn.au/_literature_93770/International_Urb

an_Road_Pricing.

Schwartz, R., Dodge, J., Smith, N.A., &Etzioni, O.

(2019). Green AI.

Song, H., Liang, H., Li, H., Dai, Z., & Yun, X.

2019.Vision-based vehicle detection and counting

system using deep learning in highway scenes.

European Transport Research Review, 11(1).

Tourani, A., Soroori, S., Shahbahrami, A., Khazaee, S.,

&Akoushideh, A. 2019. A Robust Vehicle Detection

Approach based on Faster R-CNN Algorithm. In 2019

4th International Conference on Pattern Recognition

And Image Analysis (IPRIA).

Al-Ariny, Z., Abdelwahab, M., Fakhry, M., &Hasaneen,

E. 2020.An Efficient Vehicle Counting Method Using

Mask R-CNN. In 2020 International Conference on

Innovative Trends In Communication And Computer

Engineering (ITCE).

Dorbe, N., Jaundalders, A., Kadikis, R., &Nesenbergs, K.

2018. FCN and LSTM Based Computer Vision

System for Recognition of Vehicle Type, License

Plate Number, and Registration Country. Automatic

Control and Computer Sciences, 52(2), 146–154.

Clements, L. M., Kockelman, K. M., & Alexander, W.

2020. Technologies for congestion pricing. Research

in Transportation Economics, 100863.

Wojke, N., Bewley, A., & Paulus, D. 2017. Simple online

and realtime tracking with a deep association metric.

In 2017 IEEE International Conference On Image

Processing (ICIP).

Silva S.M., Jung C.R. 2018 License Plate Detection and

Recognition in Unconstrained Scenarios. In: Ferrari

V., Hebert M., Sminchisescu C., Weiss Y. (eds)

Computer Vision – ECCV 2018. ECCV 2018. Lecture

Notes in Computer Science, vol 11216. Springer,

Cham. .

Chan, A., &Vasconcelos, N. Probabilistic Kernels for the

Classification of Auto-Regressive Visual Processes. In

2005 IEEE Computer Society Conference on

Computer Vision and Pattern Recognition (CVPR'05).

Wang, H., Lou, X., Cai, Y., Li, Y., & Chen, L. 2019. Real-

Time Vehicle Detection Algorithm Based on Vision

and Lidar Point Cloud Fusion. Journal of Sensors,

2019, 1–9.

Esther Swarna. 2018. English Typed Alphabets and

Numbers Dataset, Version 1. https://www.kaggle.com/

passionoflife/english-typed-alphabets-and-numbers-

dataset/version/1

Howard, A., Zhu, M., Chen, B., Kalenichenko, D., Wang,

W., Weyand, T., Andreetto, M., & Adam, H.

2017.MobileNets: Efficient Convolutional Neural

Networks for Mobile Vision Applications.

Strubell, E., Ganesh, A., & McCallum, A. 2019.Energy

and Policy Considerations for Deep Learning in NLP.

García-Martín, E., Rodrigues, C., Riley, G., &Grahn, H.

2019.Estimation of energy consumption in machine

learning. Journal of Parallel and Distributed

Computing, 134, 75-88.

GREE-COCO: Green Artiﬁcial Intelligence Powered Cost Pricing Models for Congestion Control

923