Journal of AI-Assisted Scientific Discovery

By Science Academic Press, USA 86

Machine Learning for Autonomous Vehicle Fleet Management and

Optimization

By Dr. Serkan Gürsoy

Associate Professor of Electrical and Electronics Engineering, Istanbul Technical University (ITU)

1. Introduction to Autonomous Vehicles

[1] A smart city integrates ICT and communication technologies into urban infrastructure to
improve the quality of life for citizens. Autonomous cars are leading components of smart
cities, and they represent the future of transportation. Urban fleet management is considered
to be an efficient method for the deployment of autonomous vehicles in smart cities, including
enterprise-level Fleet Management Systems (FMS) and consumer vehicle-sharing platforms.
Such aspects show the potential of machine learning (ML) and artificial intelligence (AI)
methods to dynamically optimize the operation of vehicles connected to the Internet of Things
(IoT). It is the main challenge to coordinate the operations of autonomous vehicles
participating in a fleet to achieve an optimal operation zone between picked up location and
requested drop-off location, and meet the specific mobile metrics (∀i λi(si, ti) < T) related to
passenger-related comfort levels for each passenger pick up etc.[2] This chapter represents a
comprehensive review related to the use of ML and AI algorithms and systems developed for
an autonomous driving system. In particular, the discussion in this chapter is also about
researched developments in the deep learning domain for a vehicle driving scenario.
Moreover, the discussion also includes research based on efficient switches and methods
suggested to keep the performance optimal while keeping the route calculation energy
efficient. In summary, in this chapter the authors discuss the future challenges and the
proposed solutions, which are important for developing autonomous driving as well as smart
transportation systems.

1.1. Definition and Types of Autonomous Vehicles

Automated driving systems work to have a nearly total absence of a human driver. The main
goal of the highly automated vehicle (HAV) in east and advanced vehicle driver assistance
systems (AVDAS) is to ensure the least human interference during emergency-driving

Journal of AI-Assisted Scientific Discovery

Volume 3 Issue 2
Semi Annual Edition | July - Dec, 2023
This work is licensed under CC BY-NC-SA 4.0.
Journal of AI-Assisted Scientific Discovery
By Science Academic Press, USA 87

situations. The researchers in some cases ignore features related to characteristics that are
similar to a human driver in the engineering of self-driving or autonomously driven vehicles.
However, the most important characteristic of human drivers in the decision-making process
is determining whether the situation is predictable or not. Although it is difficult to determine
the situation in most emergency-driving situations of the HAV due to the very low and static
visual data types, optimization factors can be constructed for the decision-making process of
the HAV. Such factors can be conveyed to the system in a decision-making process using a
fuzzy inference system, making machine learning methods’ processes much easier, faster, and
providing retainment of the system as a black box system [3].

An autonomous vehicle makes decisions based on immediate information retrieved from its
environment and long-term predictions or history. It follows a rule set and employs control
aspects to arrive at a final decision. Because of the significant steps taken in the field of
Machine Learning (ML) in recent years, it is gaining rapid attention in the field of autonomous
vehicle control [4]. In particular, ML techniques are employed in handling the decision-
making aspect focusing on two different aspects. The first formulates and solves the decision-
making pipeline using ML capacity and the second involves the direct use of ML to make
decisions. The former is employed when decision making depends on non-observable states,
when the system under consideration is dynamic, and when sporadic outcomes occur. The
latter is employed when observability is not a concern in the decision process and the machine
learning-based decisions can directly influence the outcome of the decision-making process.
Specific to autonomous vehicles, the engineering community has employed ML approaches
as a significant solution technique that takes care of many hurdles in control.

1.2. Current State and Trends in Autonomous Vehicle Technology

In the past two years, the popularity and development of smartphones, mobile Internet and
sharing economy have stimulated the rapid development of the online car-hailing industry,
and the demand for intelligent driving brings an effective potential customer group to
autonomous driving with its security and safety [5]. Hence, the development of autonomous
driving and the commercial operation of autonomous vehicle fleets have been seriously
researched by a lot of leading technology companies throughout the country. Base on the
usage scenario and market demand of driverless cars, the researches are mainly focusing on
the following aspects: the application of autonomous driving technology under the specific

Journal of AI-Assisted Scientific Discovery

Volume 3 Issue 2
Semi Annual Edition | July - Dec, 2023
This work is licensed under CC BY-NC-SA 4.0.
Journal of AI-Assisted Scientific Discovery
By Science Academic Press, USA 88

conditions of the Zijinshan Road and Harbor hospitals; The survey instrument for unmanned
driving; Path planning algorithm for unmanned vehicles.

Autonomous driving technology has been available to the public since the early years of this
millennium [4]. From the first fully autonomous vehicle in 2009 and the first grand
autonomous driving of 2013, to Google’s proposal, autonomous vehicles have gone through
many rapid changes out of the public’s sight. At present, when it comes to autonomous
vehicles, the top few companies that come to mind include the atmapy brigade, Waymo,
Baidu, Didi, and Tesla, which can be said to be the mainstream technology companies in the
field of autonomous driving. Taking Waymo as an example, It proposed and built a
completely unmanned taxi network as early as 2019, making Yosemite the first place in the
world to formally launch a commercial autonomous vehicle operation.

2. Fleet Management Challenges

The complex dependency of the parameters of the fleet and regions makes the description of
the problem difficult, what in turn influences its division into sub-tasks. Currently, there are
no publications that would directly focus on the issue of fleet management using machine
learning tools. To address this gap in the literature, we present a novel approach of fleet
management using machine learning algorithms. We abstract the fleet management problem
to data-driven decision making, to be implemented in the problem of autonomous vehicle
fleet management and dispatching on the example of a car-sharing system. The chosen fleet
management approach and vehicle operation constrained by regulation of a chosen city, allow
us to bring together the vehicle’s artificial intelligence decision making, vehicle routing
problem and impact estimation of service policies on the decision-making strategies, ride-
hailing at the customer’s side, car parking, charging and vehicle maintenance.

Machine learning methods are widely applied in on-board vehicle control tasks in level 4 and
5 autonomous vehicles (AVs) [1], in command and control of AV fleet [6], and for an impact
assessment of network composition elements on shared autonomous vehicles (SAVs) [7].
Generally, while the main focus is on managing traffic at the level of one vehicle, there are no
significant publications on the use of machine learning for fleet management connected with
the car rental. The main aim of this paper is to present a generic machine learning centric
approach to fleet management and optimization.

Journal of AI-Assisted Scientific Discovery

Volume 3 Issue 2
Semi Annual Edition | July - Dec, 2023
This work is licensed under CC BY-NC-SA 4.0.
Journal of AI-Assisted Scientific Discovery
By Science Academic Press, USA 89

2.1. Capacity Planning and Utilization

In the context of parking search, a model anticipates a driving agent who will need to find a
suitable parking location [8]. Anticipating the spatial locations of spots they will accept allows
a model to produce a distribution over driving agent behavior, where they choose actions bA
at any spatial location of consideration, a distribution over plans pA(bA), that can be used in
a model-based reinforcement learning paradigm to simulate possible behavior and calculate
a set of expected values for navigation and parking completion at those locations. Initial
investigations thus focus on situations where a driving agent is instructed where they will
need to park (not necessarily where they will end up). These distributions can be created using
a verified behavioral psychological model, in our case the model is based in part on the
Corporate Average Fuel Economy (CAFE) 2017 survey of US driver parking habits and was
verified on survey data and novel street interaction tasks in a previous work, or using k-means
clustering on street parking availability dataset.

ML-based techniques for autonomous driving such as CNN models, LSTMs, and CPTs have
been quite effective for vehicle detection, identification, localization, and tracking [9]. Recent
research in federated learning, a tailored distributed machine learning paradigm, has shown
promise in circumstances where models can be fit to a plurality of local datasets at edge nodes
rather than centralized servers. While effective in its use-case, there are several opportunities
to improve urban shared autonomous vehicle fleet management. The need exists to create
efficient last-mile strategies, design parking search and information sharing strategies, and
observe models in fleet management settings [10]. Particularly interesting is the need to better
understand how individual behavior, either of other users or between individuals, may
partially guide model results in both a parking search and a last-mile situation. This creates
opportunities where partially supervised learning or conflicting reinforcement learning may
be useful in describing anticipated behaviors in the simulation of parking navigation and
choice of shared autonomous connections.

2.2. Route Optimization and Scheduling

This section deals with decision-support methods needed to effectively optimize the path of
the vehicles from two perspectives – individual optimization, which focuses on the selection
of the best path for a single vehicle, and joint optimization, which investigates how the path
of each vehicle affects the overall traffic [11]. Consequently, these methodologies can improve

Journal of AI-Assisted Scientific Discovery

Volume 3 Issue 2
Semi Annual Edition | July - Dec, 2023
This work is licensed under CC BY-NC-SA 4.0.
Journal of AI-Assisted Scientific Discovery
By Science Academic Press, USA 90

the operational efficiency of the fleet of autonomous vehicles (route optimization, energy
management, parking management) in the daily operational activities. Our survey can be
divided approximately in three categories: (a) static and deterministic models with traditional
heuristics and machine learning techniques; (b) static and deterministic models by using
newer deep learning technique; and (c) dynamic and stochastic models realised using
reinforcement learning algorithms.

Urban traffic generates several challenges, including congestion, increased travel times, high
energy consumption, and carbon dioxide emissions. It is crucial to mitigate these issues
through new urban mobility paradigms, such as ITS, VANET, electric vehicles, and
Autonomous Vehicle technology [12]. These issues can be mitigated through the integration
of intelligent transportation systems (ITS), future generation VANETs [13]. These challenges
present a new and prominent research areas in advanced automation techniques (hybrid AI
+ Data-driven models, reinforcement learning, etc.) to support decision-making process of
autonomous vehicles.

3. Machine Learning Basics

Machine Learning (ML) and AI algorithms are required to be implemented, tested, and
evaluated on the basis of current and future extensive data coming from control units, sensors
typical of autonomous vehicle technology, Internet of Things (IoT) devices, social networks
and many other sources of data . Herein, the implementation, the commissioning, and the
testing may be challenging since possible changes in vehicles performances connected to
software driven performance failures shall be promptly recognized and corrected. Therefore,
the final aim of this note is the introduction and description of the most recent and successful
applications of AI and ML to automate a series of deceptive last several passengers of alerts
stemming from a Vehicle Health Monitoring sub-system.

article_main_idea: There is a growing interest in developing and studying methods for

Commercial Connected Automated Mobility (CCAM) concept, in which mobility services are
provided by commercial companies possessing fully or partially connected automated
vehicles . In , the authors introduced the multi-agent Discrete Event Systems based paradigm
for “Connected Automated Vehicle Fleets”. In that model, a permanent bilateral information
flow is established between the AV (A) and a remote Central Management (CM) unit, while
the CM unit acts as an intermediate authority (total traffic manager) that directly supervises

Journal of AI-Assisted Scientific Discovery

Volume 3 Issue 2
Semi Annual Edition | July - Dec, 2023
This work is licensed under CC BY-NC-SA 4.0.
Journal of AI-Assisted Scientific Discovery
By Science Academic Press, USA 91

vehicles functionalities in terms of activation and deactivation of the automated level of the
vehicles, handling of the car sharing and rental operations and performs safety supervising
functions.

3.1. Supervised, Unsupervised, and Reinforcement Learning

The use of machine learning (ML) techniques in the development and optimization of
autonomous vehicle-fleet management systems is gaining in popularity. A review of the
recent state of the art indicates that a majority of the techniques abstraction levels range from
preplanning (off-line) to online (perception and decision-making), which leads to a very
centralized, monolithic, and myopic approach. Another trend observed in the review is that a
large number of the perceivers are still—as of this review writing—handcrafted feature
extractions or drivable area detection pipelines where deep learning based on segmentation
(supervised or unsupervised) is the most popular approach, although some real-world
autonomy platforms like Tesla’s Autopilot [14], Waymo, Cruise, Aurora, Zoox, and Argo are
also investing effort to convert perception problems into supervised machine learning by
focusing on realistic high performance edge deployment with efficient online learning.
However, in dealing with the challenges of optimising fleet management for ceteris paribus
scenarios, such as demand forecasting, routing, and optimizing of dis/init low-planning
practical resolutions, only simple models like linear regressions or based approaches on
clustering (k-means, etc.) are applied. Furthermore, the existing review is not explicitly
focussed on scalable, promising novel methods that are being applied to fleet management
problems in the literature. The most popular deep learning techniques for automatic feature
extraction and high-precision feature map labeling are supervised, unsupervised, and
reinforcement learning. Convolutional neural network (CNN) based segmentation, like, U-
Net, SegNet, and FCN, is the most popular implementations for supervised learning because
they are robust and easily trained on large quantities of data. Recent advancement in
unsupervised learning for scene understanding and neglecting manual pixel-level annotation
in semantic segmentation, has led to a flood of techniques for simultaneous per-controlling
slot operations of cameras in congested areas and to avoid infrastructure investments on
optical sensors, such as lidar, by using passive sensor input like a monocular camera [15].
‘Reinforcement learning’ can use data without any labeled ground truth data and train models
through end-to-end control. However, there are some among researchers who postulate that
there’s little support for substantial benefits of integrated curriculum learning in autonomous

Journal of AI-Assisted Scientific Discovery

Volume 3 Issue 2
Semi Annual Edition | July - Dec, 2023
This work is licensed under CC BY-NC-SA 4.0.
Journal of AI-Assisted Scientific Discovery
By Science Academic Press, USA 92

driving [16]. Compared to other machine learning techniques, k- means, DBSCAN,

hierarchical, and spectral clustering are less utilized for classification, regression and route-
based discrete performance optimization problems. The reality is that these unsupervised
learning algorithms are not always consistent when they are forced to segment sequential
waypoint plan arrivals or large datasets and they are not implemented to seek minimum
scalar benefit for fleet and demand satisfaction on metamer prickpea ata.

4. Machine Learning Applications in Fleet Management

Given continuous car usage and the spread of Internet of Things and wireless sensor
networks, traffic congestion is a raising problem in many of today’s major cities.
Simultaneously, having efficient traffic load and distribution helps to extend battery lifetime
of the cars and provides proper movement of urgent service vehicles such as ambulances and
fire trucks. Many solutions are suggested to reduce the traffic congestion. Such solutions are
either as simple as changing the traffic light duration or as complicated as routing vehicles by
considering different traffic parameters in real time [2]. Developing Coordination protocols
that react to the traffic condition in real time have also been proposed to better manage the
traffic. Even more advanced solutions like utilizing flying cars have been suggested to reduce
traffic congestion. However, there is one thing that all these solutions have in common. They
all require precise information about the driver's environment to make clear decisions. This
need for detailed information can be met by proper usage of machine learning methods that
can help to predict future traffic patterns and to analyze vehicles routing history by taking
into account different traffic parameters [17].

In the last several decades, intelligent transportation systems (ITS) have been focused on
mitigating the negative effects of transportation on the environment. Given the importance of
energy consumption, many approaches have been adopted to improve-traffic operation, such
as flying cars [18]. These cars have proven to cause fewer accidents and reduce latency in city
modernization. Moreover, several routing algorithms have been proposed in order to reduce
the consumption and emissions of cars in dynamic environments.

4.1. Demand Prediction and Forecasting

[19] Autonomous vehicle fleet management and optimization requires accurate demand
forecasting. To optimize consumer interests and asset utilization, and to reduce operating

Journal of AI-Assisted Scientific Discovery

Volume 3 Issue 2
Semi Annual Edition | July - Dec, 2023
This work is licensed under CC BY-NC-SA 4.0.
Journal of AI-Assisted Scientific Discovery
By Science Academic Press, USA 93

costs, fleet service providers need accurate forecasts of passenger demand. The purpose of
Demand Forecasting is to predict consumer ride requests in the upcoming period for different
fleet areas, levels of traffic congestion, and time of the day. Based on the predicted future and
current consumer demand and incomplete information, the planning strategy for fleet
vehicles in the next period can be optimized, and the results can be used for deployment of
fleets and optimal use of the available vehicle-sharing pool.[20] Demand prediction and
forecasting is an important part of vehicle dispatching. Through the demand prediction and
forecasting of the fleet service area, the future demand of passengers can be effectively
predicted, and the passenger demand situation of the entire service area in the future can be
better understood, and the efficiency and quality of the dispatch of the fleet can be greatly
improved. A variety of demand prediction and forecasting methods have been used in fleet
dispatch management. Traditional traffic forecasts are mainly based on classical time series
analysis and regression methods. However, these methods may need to consider more
influential factors, and it is not easy to adapt to various changes. This area is relatively large,
so many researchers have proposed to use machine learning or deep learning forecasting
methods to improve the efficiency and accuracy of demand forecasting and dispatching of
fleet management models.

4.2. Dynamic Pricing Strategies

In smart cities, autonomous or self-driving vehicle fleets will introduce a fleet management
problem in the supply chain, which can be handled in different ways as an optimization
problem, for example, by controlling the locations and movements of vehicles. In this paper,
a machine learning method is integrated in a fleet management decision-making model,
where the objective is to provide the most convenient autonomous vehicle fleet for customers
and for the service provider. This paper will provide a machine learning optimization tool for
private and shared self-driving transportation. For private self-driving cars, a system for
mobility on demand, with users leaving their self-driving car wherever they like, will be
developed. For shared self-driving vehicles, where passengers share a car, an earnest
passenger-first approach will be implemented. This paper describes the functioning of an
autonomous robotic vehicle, built from a Mobile-Robot K3. It provides tools for developing a
dedicated decisional model specifically designed for that application. We describe an
architecture based on finite state machines (FSM) concepts. Different scenarios in decision
theory can be distinguished. They differ by the objective, the domain on which the decision is

Journal of AI-Assisted Scientific Discovery

Volume 3 Issue 2
Semi Annual Edition | July - Dec, 2023
This work is licensed under CC BY-NC-SA 4.0.
Journal of AI-Assisted Scientific Discovery
By Science Academic Press, USA 94

made and the decision makers involved. The selected decision-making model is often ruled
by technical constraints of the robotics systems. In the model proposed in this paper, a
hierarchical FSM decision model has been used. The first step in this model aims at classifying
the contextual scenario. The next steps consist in identifying the “best” action in the given
scenario, according to one or more strategies associated to the current behavior.

5. Optimization Techniques

Prediction-based learning allows us to minimize prediction error, characteristic-based

learning allows preventive schedule changes, and exploiting the model to optimize is model-
based learning. The learning models we describe in this section are the most commonly used
and more popular among other models in literature. [21] The predictionbased learning,
mainly used to model relations between energy consumption and speed information, are
various machine learning models, such as simple linear regression, multivariate kernel
regression, and Gaussian process regression. Data-driven clustering has been applied to
classify the service areas according to the conditional probability distribution of the future
customer requests (classification-based learning). A model estimating energy cost and the
battery charge probability was developed using reinforcement learning, which then
optimized the driving policy of electric vehicles. Algorithms were proposed in the two-stage
scenario: a machine learning model predicts the weather-based energy costs, and then this
result is integrated in the optimization formulation. [22] In the multi-crieria optimization
branch, there is simultaneous reversal between the prediction and optimization problems,
where regression-based learning is the most used one among others. The machine learning
tools applied to the location routing problem in intuition and augmented multi-objective
optimization include the elastic net, least angle regression, ridge regression, and t-distributed
stochastic neighbor embedding (t-SNE). Constraint handling techniques that exploit the
correlation between the objectives and constraints coordinated with binary classification,
binary relevance, and binary relevance weighted with a margin scaling model are used to
implicitly predict and optimize simultaneously. The complex scheduling problem is modeled
as a set of regression, classification, and ranking problems. Epsilon support vector regression,
ordinal regression, Gaussian process regression, and the support vector data description are
popular among the optimization-based learning methods. Providing a disappointingly
meager understanding of the combined effect of all the criteria on the world may unlock new
insights.

Journal of AI-Assisted Scientific Discovery

Volume 3 Issue 2
Semi Annual Edition | July - Dec, 2023
This work is licensed under CC BY-NC-SA 4.0.
Journal of AI-Assisted Scientific Discovery
By Science Academic Press, USA 95

resulting optimization problem comprises coordination and scheduling over all vehicles
based on the future predictions. The challenge is to maintain some reserve capacity in resource
management to satisfy a certain performance level in the presence of uncertainties.
Regridding and recalculating accurately is not possible in a “real time” or “on the fly”
optimization environment. However, an adequate structure equipped with learningbased
methods can provide flexibility in prediction and optimization. While there are learning
techniques that are predictions-based, characteristic-based, or model-based learning, we
develop models that combine various approaches where applicable.

[23] Optimization is essential for efficient autonomous vehicle fleet management. It includes
tasks such as predicting future states and customer requests with high precision. The

5.1. Linear and Integer Programming

These predictions are fed into an IP-based vehicle routing problem (VRP) model, typically
solving it real-time. When solving IPs for real-time AV decisions, it is crucial to tailor the ILP
model to the considered application— the classic ILP fundamentals are insufficient. Some AV
ILP strategies : directly model the RL problem as a MIP. However, usually we include tight
training-time generated upper bounds on the maximal number of passengers to serve and
vehicles to use, since the MIP is prone to degenerate solutions that result in slow solution
times and extensive iterations . Acknowledging the LP relaxation of the MIP, another key
development here is that reinforcement learning can guide policy improvement in this
relaxation, by biasing LP variable bounds and therefore the decomposition’s subnet
structures.

A linear program (LP) aims to optimize a linear objective function over a constraint set that
can be naturally represented using matrices of constants and variables. Integer programming
(IP) extends LP to optimize over not just reals, but also integers. Therefore, IPs can represent
all combinatorial optimization problems up to some imprecisions [24]. Hence, IPs have broad
applications in urban traffic management and MAFM of goods, services, and people [25]. Yet
their only-memory approach to linearize nonlinear and disjunctive constraints can be quite
conservative and prune a large fraction of feasible and potentially high-quality solutions. With
growing availability of data and computing resources, tailored algorithms that leverage this
to optimize IP solutions are desirable, notably in the context of AVs. Within this context,

Journal of AI-Assisted Scientific Discovery

Volume 3 Issue 2
Semi Annual Edition | July - Dec, 2023
This work is licensed under CC BY-NC-SA 4.0.
Journal of AI-Assisted Scientific Discovery
By Science Academic Press, USA 96

random forests and machine learning (ML) models feature prominently in the recent literature
[23]. For instance, they are used to predict future vehicle uses.

5.2. Genetic Algorithms and Simulated Annealing

\newline The implementation of a ride-hailing solution with Genetic Algorithms (GAs) and
Simulated Annealing (SA) is already discussed in this section. A ride-hailing service like Uber
in a city with huge number of Autonomous Vehicles (AVs) would be seen in the future.
Therefore, fleet management, which needs the scheduling of vehicle tasks and their trajectory
control, is particularly significant for the system to achieve riding quality and high efficiency
at the same time. [26] Beyond traditional methods, researchers should also explore deep
learning-based strategies and low-latency real-time optimization methods, including
applying the methods highlighted in this section.

[27] The solution to the routing problems of managing ride-hailing vehicles or autonomous
vehicle (AV) fleets is called the vehicle rerouting problem. Genetic algorithms (GAs) can be
used to fix vehicle trajectories at the departure times of passengers to maintain an optimized
plan while Simulated Annealing (SA) calculates a probabilistic turn probability for rerouting.
GAs and SA, MAS, Reinforcement Learning, meta-heuristic & traditional methods considered
for solving ride-hailing vehicle management regarding minimizing Distance Traveled,
Frustration, and Number of Passenger Requests Refused problems.

6. Simulation and Testing

Before modeling the autonomous control/control strategies of an entity, the entities

themselves (henceforth, vehicles) should be modeled in a straightforwardly observable way,
and connected to the rest of the vehicle-control infrastructure using standard vehicle- control
interfaces. This ensures that control strategies developed using only high-level abstract inputs
(like “desired throttle position” for a passenger vehicle under cruise control) can be directly
deployed to the real entity. This is because it is common that a software control strategy
outperforms the original goals of that entity. For example, a dynamic traffic management
system allows passengers to substantially increase their willingness to accept longer trip times
if they know that the vehicles they are in are operated by akin dynamic management
strategies. Such a transition in vehicle operation can only take place if the developers of the
dynamic management strategy can show that the new strategy doesn’t exacerbate the gap

Journal of AI-Assisted Scientific Discovery

Volume 3 Issue 2
Semi Annual Edition | July - Dec, 2023
This work is licensed under CC BY-NC-SA 4.0.
Journal of AI-Assisted Scientific Discovery
By Science Academic Press, USA 97

between the desired behavior of the controlled vehicle (for example, surge speed) and the as-
built or the as-driven vehicle. The damage control can take the form of additional controlled
entities modeling layer as has been modeled through careful choice of constrained control
inputs managed by a particular ECU modeling the acquired constraints for critical entities
and the use of zero-shot imitation learning via zero-shot action-conditional RNN for non-
critical entities [28].

[29] [30]Simulating autonomous vehicles in the real world is costly and can be unsafe.
Simulations are a safer and run under more controlled conditions than real-world
testing~\cite{DBLP:journals/corr/abs-2008-11335}. Open-source software like Microsoft
AirSim, Ford’s CARLA, and even NASA World Wind with Gazebo can model both the
environment in which the vehicle drives and the subsequent vehicle behavior. This modeling
ranges from modeling the physics and vehicle controls to modeling every individual
pedestrian in the environment. The simulator can also introduce multiple types of noise,
distractions, and other driving hindrances to provide more robust testing than real-world
testing can. As a result, current research and development in autonomous vehicle technology
emphasize a testing-first approach.

6.1. Virtual Environments and Simulators

Open-source traffic simulation packages, by focusing on quick manual data visualization and
explicit user dynamic interactions, typically offer real-time, easily adjustable, interactive
traffic scenarios. They require only limited physical domain knowledge or programming
skills and are aimed towards the development of a large, loosely coupled scientific
community. When it comes to reversible planning, model training and data validation are
needed. The control problems with an AD and with an AV fleet require an extensive amount
of data and extensive real-world testing and validation. This is due to the diversity of
essentially random and chaotic traffic interactions which involve many entities with different
behaviors and control capabilities. In these situations, the safe, easy to repeat and controllable
scenarios provided by V&E&S are potentially a perfect training, validation, exploration and
AI traffic engineering choice.

The machine learning approach necessitates extensive training and validation data coming
from real traffic flows, which are oftentimes not readily available, are difficult to collect, and
introduce safety and privacy constraints [31]. Virtual environments and simulators (V&E&S)

Journal of AI-Assisted Scientific Discovery

Volume 3 Issue 2
Semi Annual Edition | July - Dec, 2023
This work is licensed under CC BY-NC-SA 4.0.
Journal of AI-Assisted Scientific Discovery
By Science Academic Press, USA 98

feature the role of fully transparent, adjustable, and re-creatable traffic system representations
useful for safe artificial intelligence (AI) algorithm training and for controlled AV-FM
(autonomous vehicle fleet management) scenario testing, also denoted as off-line and on-line
training and validation scenarios [28]. Simulators allow for complete control over all entities
of the simulated environment, which opens up and provides the feasibility of strictly targeted
model exploration and performance evaluation, in bounded and unbounded AD
(autonomous driving) and AV-FM training and validation scenarios. Most of the ensemble-
based, open-source simulators personalize either physics-based simulations engine, traffic
flow and congestion complexity, API (application programming interfaces), or the ease of use,
but in most cases they offer only limited user interactions [30].

7. Case Studies and Real-World Applications

In this paper, we propose a novel effective model, called LaSSLSTM, in order to better predict
other agents’ behavior in realtime. It is crucial for any autonomous driving agent to be able to
predict what will happen next on a road in the near future. Curb weight, vehicle’s movements,
and vehicle’s existence on the road could be obtained from raw data by applying initial feature
extraction. Then, the trajectories representing the future path of agents are determined by
using predictions. In order to understand that LaSS-LSTM is better in end-to-end highway
driving scenarios than lane-aware LSTM and plain LSTM, LaSS-LSTM should be analyzed
and compared in terms of prediction success, following traffic rules, staying in its lanes, and
differentiating vulnerable users.

[16] Machine learning is playing a critical role in helping autonomous and connected vehicles
react to other road users, traffic, and infrastructure. This paper summarized our efforts with
advanced machine learning technologies for autonomous applications, including vehicle
trajectory prediction, autonomous navigation, and autonomous fleet management and
optimization, which are critical for the focal field. Overall, the knowledge on vehicular data
analysis with machine and deep learning in this paper implements a system that records,
analyses, and provides real-time feedback. This system uses machine and deep learning
models, which are trained with driving behavior data using 18 different features such as
speed, acceleration, etc. for each driver segment using supervised learning. After model
training, by using the physical hardware setup, the driving behavior of a driver can be
evaluated based on the incoming data. Furthermore, this system uses ML models to find

Journal of AI-Assisted Scientific Discovery

Volume 3 Issue 2
Semi Annual Edition | July - Dec, 2023
This work is licensed under CC BY-NC-SA 4.0.
Journal of AI-Assisted Scientific Discovery
By Science Academic Press, USA 99

driving changes of the drivers and allows to interact with an environment’s dynamic
components like other cars, cyclists, and pedestrians. This system aims to generate immediate
safety and eco-friendly behavioral warnings for drivers regardless of the restriction.[32]
Autonomous vehicles running in the real world will sooner or later need reliable driving
models on other road users, including the vehicles to predict the position and behavior of
surrounding vehicles in the near future. To this end, we proposed a novel predictor called
lane stream attention-based LSTM (LaSSLSTM), which effectively leverages road information
to improve prediction. Our proposal builds on the assumption that public driving datasets
are generally collected under the similar lane configurations and prioritize stream information
from the lanes to improve the accuracy of trajectory prediction in challenging scenarios. In
highway settings, we proposed a method that can improve the prediction capability of
trajectory prediction models when there are not enough historical data.

7.1. Uber and Lyft: Dynamic Pricing and Route Optimization

In the following, I present some online dynamic routing and pricing strategies. presents a
selection of algorithms that efficiently and successfully control the car routing and fleet
management in the context of autonomous electric vehicle on-demand ride-sharing services
with substantial battery recharging constraints, demonstrating an average gain in number of
served customers per day up to 40% or more in our experiments. leans on carpooling
transportation features through a hierarchical multi-agent approach which enables effective
fleet operation and provides sustainable transportation alternatives ensuring convenient,
cost-effective, and demand-responsive rides. The flexibility and control of the entire ride tours
respecting the market conditions of carpooling allows stakeholders to partially shape their
desired travel pattern over time and space, while the adoption of a reinforcement learning
algorithm provides an efficient probabilistic approximation to the TAR pool demand function
in situations where the space of states and demand pattern are unknown and/or large.

[33] [34]Operational and decision-making strategies to manage fleets of autonomous vehicles

require considering dynamic routing and vehicle dispatching, as well as the real-time
decisions to take, such as dynamic pricing and trip matching. This section aims at reviewing
the strategic and operational challenges faced by ride-hailing companies' fleet optimization
systems. As to be expected, Lyft and Uber are two of the most requested and well-known
mobility service platforms within urban areas. Lyft and Uber's systems operate in a similar

Journal of AI-Assisted Scientific Discovery

Volume 3 Issue 2
Semi Annual Edition | July - Dec, 2023
This work is licensed under CC BY-NC-SA 4.0.
Journal of AI-Assisted Scientific Discovery
By Science Academic Press, USA 100

yet different operational model which is further described below. This section analyzes both
companies with the dual aim of identifying strategic and operational differences and
providing an overview of the state-of-the-art methods needed to address their issues.

8. Ethical and Legal Considerations

Considering the whole range of challenges that we are facing, including technical, operational
and people’s behavior challenges this research in particular: 1) focused on the development
and validation of Machines Learning (ML) algorithms for covering both the short and long
term multimodal predictive optimization of autonomous vehicles fleet management features,
2) definitions the application of ML-based approaches for Ethical and Legal considerations in
the manage and operations of AVMs. Indeed, the significant upgrade in AV safety capabilities
that the Industry 4.0 new approach to risk analysis, reduction and management should
actively manage the Ethical & Legal issues related to the new risk factor must be properly
considered. Adopting the above mentioned parallel approach to the evitable reduction in the
residual risk (namely by the inclusion of new safety appropriate hardware and/or performing
on-road extensive verification and validation tests), proper the serious of ethical and legal
accidents avoidance system on on of these vehicles can provide a form of very well welcomed
warranty to the public. [35]

Enabling technologies and spiking consumer interest have become the force driver for the
Smart Mobility/Transportation Services, including the burgeoning deployment of
Autonomous Vehicles (AV). Autonomous vehicle (AV) will transform transportation by
doing their work more efficiently, saving lives, and mitigating negative environmental
impacts [36]. The future of AV remains undefined as concerns exist amongst the public and
regulators with the potential harmful consequences. As pointed out by Massaro and Taddia,
the development and deployment of fully automated vehicles raises ethical concerns,
especially in extreme traffic situations (‘trolley problem’) where they may need to make
challenging ethical choices [37]. In today's AV market however, ethical concern is a marginal
issue, solving different technological, legal, or technical barriers is rather the focus of the
industry. However, these challenges deserve addressing by including ethical and legal criteria
in the development of mobility solutions, both in the software or hardware system and
regulatory fields.

8.1. Data Privacy and Security

Journal of AI-Assisted Scientific Discovery

Volume 3 Issue 2
Semi Annual Edition | July - Dec, 2023
This work is licensed under CC BY-NC-SA 4.0.
Journal of AI-Assisted Scientific Discovery
By Science Academic Press, USA 101

Guideline 6!seudo-logical approaches are becoming a goal driven venue for machine learning
approaches to be implemented. Therefore the requirement of verification and validation must
be ensured that the model is safety, dependable and has been built through a security risk
modeling life cycle development and testing. This means that foundational dumb data AI,
deep learning AI and safety and security AI models should be trained on dirty and
confounding datasets. Dirty data to ensure the baselines are the minimums for ML/AI models
in associated self-driving cars sensors and map localization understanding. Confounding data
to further train deeper models that are more effective in a larger set of tasks.leta models like
SWAGEN or Adversarial NLG outperforms foundation AI models for generating multi turn
dialogues and Class discriminative models. In case we require foundation models with higher
granularity, then AGPT-4 can be utilized to convert model from AGPT-2B model to AGPT-3B
model by either using the transformer knowledge or distilled model method.

Predictable action has been achieved in implementing Single Net classifiers, where a data
generating model is trained to generate favorable trajectories and replayed to train a single
policy network. On the other dimension, Reinforcement Learning (Sutton 1998) technique, has
also been shown to perform even better while training the learning model to take multiple
number of steps in the future for each training data. Once a physical network has been trained,
it can be discretized to be used in shorter distance prediction or other experimental use-cases.
Deep learning (LeCun et al. 2015, Goodfellow 2016, He et al. 2015) can be a good choice in
many industry-level data-driven production systems, machine vision, and natural language
processing especially when the input data is images or a high figure representation language
like natural language. Though other type of neural networking would suit good, the deep
learning however with the associated convolution neural network (CNN) has outperformed
in many learning models performance especially in image recognition and object detection.
Nevertheless, it should be noted that machine learning model of autonomous vehicle has
entered into a state, where acting technically without causing risk.

Several limitations: static and dynamic placement optimization models for MECs are scientific
preliminary research; much effort should be devoted to confronting the NP-hard nature of the
problem and developing effective heuristic or meta-heuristic algorithms; real-world data from
autonomous vehicles is hard to collect and are often confidential. autonomou driving data can
be available through simulators. The bulk of computing must be performed at the MEC on a
real-time basis. The need to carefully manage distributed data processing must be integrated

Journal of AI-Assisted Scientific Discovery

Volume 3 Issue 2
Semi Annual Edition | July - Dec, 2023
This work is licensed under CC BY-NC-SA 4.0.
Journal of AI-Assisted Scientific Discovery
By Science Academic Press, USA 102

into autonomous driving strategies. This is the driving force behind the use of emergent
technologies such as edge computing, federated learning, and true hybrid learning with
machine, distributed ETL and Mec. Model explainability in autonomous driving system also
plays a critical role to solve a following challenge. An explainable model shows the certain
evidence or explanations to the reason behind the prediction result. Ran’ lot “” has shown that
a lot of theoretical relation and techniques would be mentioned if a machine learning model
is selected for an Autonomous Vehicle (AV). These machine learning models in essence is a
function that takes the input state of the surrounding neighborhood and predicts the good
maneuver in order to transit from the current state to a desirable state.

Managing autonomous vehicle fleets from a central server requires real-time data stream
synchronization including PALD (passenger autonomous driving) data and OAD (online
autonomous driving) training data. Model explainability requires that experiments be
designed and analyzed so that performance can be understood in the context of the
underlying physics. For the components of AMoD, the CCN model with graph-convolutional
networks is promising. Foundation models such as LLMs/VLMs may have potential
performance and accuracy advantages, but they have challenges that can make them worse
than the best current approaches when used as previously proposed.

9. Conclusion and Future Directions

The benefits of using machine learning for predicting the Electric Vehicles activity beyond the
capabilities may become organising smarter and more efficient public transport systems and
in this work the first results obtained combining machine learning clustering model and
queuing are represented in a closed-loop city route scenario. The effective urgency of
obtaining real-world data and creating prediction models has not allowed justifying the real
achievement of effective and perceptible benefits on transit service performances. Now, the
future applications in which the here represented predictive model can find application are
exemplified by the real-world problems already solved by neural network prediction driven
solution. This last part provides several real-world analogous application areas already
tackled by neural network based prediction foundational improvement, to possibly anticipate
at which kind of problems currently faced by existing shuttle services the proposed model
could be efficiently and effectively used in a near-future perspective.

Journal of AI-Assisted Scientific Discovery

Volume 3 Issue 2
Semi Annual Edition | July - Dec, 2023
This work is licensed under CC BY-NC-SA 4.0.
Journal of AI-Assisted Scientific Discovery
By Science Academic Press, USA 103

[19], [32] The work presented in this chapter aims to demonstrate the potential of combined
machine learning and queueing theory models for fleet management optimization of
autonomous electric vehicles operating in mixed fleet, closed-loop shuttle circuit scenarios
within an urban environment. This computational approach is designed to provide dynamic
and optimal deployment and operation strategies, based on real-time vehicle localisation data,
electric vehicle state-of-charge data, and queueing theory based demand forecast, to suffering
transit service operators to transit agencies, with the target of enhancing the daily services
provided to the passengers while enhancing the fuel and the duration of battery provided by
operating electric autonomous vehicles that may have impacts on the entity’s environmental
performances [36].

References:

1. [1] A. Qayyum, M. Usama, J. Qadir, and A. Al-Fuqaha, "Securing Connected &

Autonomous Vehicles: Challenges Posed by Adversarial Machine Learning and The
Way Forward," 2019. [PDF]
2. [2] Y. Yang, T. Dwyer, K. Marriott, B. Jenny et al., "Tilt Map: Interactive Transitions
Between Choropleth Map, Prism Map and Bar Chart in Immersive Environments,"
2020. [PDF]
3. [3] A. Biglari and W. Tang, "A Review of Embedded Machine Learning Based on
Hardware, Application, and Sensing Scheme," 2023. ncbi.nlm.nih.gov
4. Tatineni, Sumanth. "Customer Authentication in Mobile Banking-MLOps Practices
and AI-Driven Biometric Authentication Systems." Journal of Economics & Management
Research. SRC/JESMR-266. DOI: doi. org/10.47363/JESMR/2022 (3) 201 (2022): 2-5.
5. Vemori, Vamsi. "Evolutionary Landscape of Battery Technology and its Impact on
Smart Traffic Management Systems for Electric Vehicles in Urban Environments: A
Critical Analysis." Advances in Deep Learning Techniques 1.1 (2021): 23-57.
6. Shaik, Mahammad, Srinivasan Venkataramanan, and Ashok Kumar Reddy Sadhu.
"Fortifying the Expanding Internet of Things Landscape: A Zero Trust Network
Architecture Approach for Enhanced Security and Mitigating Resource
Constraints." Journal of Science & Technology 1.1 (2020): 170-192.

Journal of AI-Assisted Scientific Discovery

Volume 3 Issue 2
Semi Annual Edition | July - Dec, 2023
This work is licensed under CC BY-NC-SA 4.0.
Journal of AI-Assisted Scientific Discovery
By Science Academic Press, USA 104

7. [7] A. Gupta, A. Shaharyar Khwaja, A. Anpalagan, L. Guan et al., "Policy-Gradient and

Actor-Critic Based State Representation Learning for Safe Driving of Autonomous
Vehicles," 2020. ncbi.nlm.nih.gov
8. [8] N. Strauß, L. Rottkamp, S. Schmoll, and M. Schubert, "Efficient Parking Search
using Shared Fleet Data," 2024. [PDF]
9. [9] A. M. Elbir, B. Soner, S. Coleri, D. Gunduz et al., "Federated Learning in Vehicular
Networks," 2020. [PDF]
10. [10] A. Arizala, A. Zubizarreta, and J. Pérez, "A Complete Framework for a Behavioral
Planner with Automated Vehicles: A Car-Sharing Fleet Relocation Approach," 2022.
ncbi.nlm.nih.gov
11. [11] S. Paul, J. Witter, and S. Chowdhury, "Graph Learning-based Fleet Scheduling for
Urban Air Mobility under Operational Constraints, Varying Demand &
Uncertainties," 2024. [PDF]
12. [12] H. Kamal, M. Picone, and M. Amoretti, "A Survey and Taxonomy of Urban Traffic
Management: Towards Vehicular Networks," 2014. [PDF]
13. [13] L. Baty, K. Jungel, P. S. Klein, A. Parmentier et al., "Combinatorial Optimization
enriched Machine Learning to solve the Dynamic Vehicle Routing Problem with Time
Windows," 2023. [PDF]
14. [14] L. Fridman, "Human-Centered Autonomous Vehicle Systems: Principles of
Effective Shared Autonomy," 2018. [PDF]
15. [15] I. H. Sarker, "AI-Based Modeling: Techniques, Applications and Research Issues
Towards Automation, Intelligent and Smart Systems," 2022. ncbi.nlm.nih.gov
16. [16] S. Mahmoud, E. Billing, H. Svensson, and S. Thill, "How to train a self-driving
vehicle: On the added value (or lack thereof) of curriculum learning and replay
buffers," 2023. ncbi.nlm.nih.gov
17. [17] H. Ye, L. Liang, G. Ye Li, J. B. Kim et al., "Machine Learning for Vehicular
Networks," 2017. [PDF]
18. [18] O. Rinchi, A. Alsharoa, I. Shatnawi, and A. Arora, "The Role of Intelligent
Transportation Systems and Artificial Intelligence in Energy Efficiency and Emission
Reduction," 2024. [PDF]
19. [19] D. K. Galloway, J. J. M. in 't Zand, J. Chenevez, L. Keek et al., "The Influence of
Stellar Spin on Ignition of Thermonuclear Runaways," 2018. [PDF]

Journal of AI-Assisted Scientific Discovery

Volume 3 Issue 2
Semi Annual Edition | July - Dec, 2023
This work is licensed under CC BY-NC-SA 4.0.
Journal of AI-Assisted Scientific Discovery
By Science Academic Press, USA 105

20. [20] H. Hu, J. Xu, M. Liu, and M. K. Lim, "Vaccine supply chain management: An
intelligent system utilizing blockchain, IoT and machine learning," 2022.
ncbi.nlm.nih.gov
21. [21] P. Agand, A. Kennedy, T. Harris, C. Bae et al., "Fuel Consumption Prediction for
a Passenger Ferry using Machine Learning and In-service Data: A Comparative
Study," 2023. [PDF]
22. [22] F. Hammes, A. Hagg, A. Asteroth, and D. Link, "Artificial Intelligence in Elite
Sports—A Narrative Review of Success Stories and Challenges," 2022.
ncbi.nlm.nih.gov
23. [23] A. Paul, M. W. Levin, S. Travis Waller, and D. Rey, "Data-driven Optimization for
Drone Delivery Service Planning with Online Demand," 2024. [PDF]
24. [24] T. H. Hejazi, Z. Ghadimkhani, and A. Borji, "A learning-based solution approach
to the application placement problem in mobile edge computing under uncertainty,"
2024. [PDF]
25. [25] T. Y. Ma and Y. Fang, "Survey of charging scheduling, fleet management, and
location planning of charging stations for electrified demand-responsive transport
systems: methodologies and recent developments," 2021. [PDF]
26. [26] Y. Jiao, X. Tang, Z. Qin, S. Li et al., "Real-world Ride-hailing Vehicle Repositioning
using Deep Reinforcement Learning," 2021. [PDF]
27. [27] I. De Vincenzo, I. Giannoccaro, and G. Carbone, "The Human Group Optimizer
(HGO): Mimicking the collective intelligence of human groups as an optimization tool
for combinatorial problems," 2016. [PDF]
28. [28] B. Poudel and W. Li, "CARL: Congestion-Aware Reinforcement Learning for
Imitation-based Perturbations in Mixed Traffic Control," 2024. [PDF]
29. [29] O. Saoudi, I. Singh, and H. Mahyar, "Autonomous Vehicles: Open-Source
Technologies, Considerations, and Development," 2022. [PDF]
30. [30] J. Felipe González-Saavedra, M. Figueroa, S. Céspedes, and S. Montejo-Sánchez,
"Survey of Cooperative Advanced Driver Assistance Systems: From a Holistic and
Systemic Vision," 2022. ncbi.nlm.nih.gov
31. [31] D. Wang, L. Gao, Z. Lan, W. Li et al., "An Intelligent Self-Driving Truck System
for Highway Transportation," 2022. ncbi.nlm.nih.gov

Journal of AI-Assisted Scientific Discovery

Volume 3 Issue 2
Semi Annual Edition | July - Dec, 2023
This work is licensed under CC BY-NC-SA 4.0.
Journal of AI-Assisted Scientific Discovery
By Science Academic Press, USA 106

32. [32] D. Yu, H. Lee, T. Kim, and S. H. Hwang, "Vehicle Trajectory Prediction with Lane
Stream Attention-Based LSTMs and Road Geometry Linearization," 2021.
ncbi.nlm.nih.gov
33. [33] E. Bokányi and A. Hannák, "Understanding Inequalities in Ride-Hailing Services
Through Simulations," 2020. ncbi.nlm.nih.gov
34. [34] J. Si, F. He, X. Lin, and X. Tang, "Vehicle Dispatching and Routing of On-Demand
Intercity Ride-Pooling Services: A Multi-Agent Hierarchical Reinforcement Learning
Approach," 2023. [PDF]
35. [35] P. Radanliev, O. Santos, A. Brandon-Jones, and A. Joinson, "Ethics and responsible
AI deployment," 2024. ncbi.nlm.nih.gov
36. [36] S. Paiva, M. Abdul Ahad, G. Tripathi, N. Feroz et al., "Enabling Technologies for
Urban Smart Mobility: Recent Trends, Opportunities and Challenges," 2021.
ncbi.nlm.nih.gov
37. [37] L. Luxmi Dhirani, N. Mukhtiar, B. Shankar Chowdhry, and T. Newe, "Ethical
Dilemmas and Privacy Issues in Emerging Technologies: A Review," 2023.
ncbi.nlm.nih.gov

Journal of AI-Assisted Scientific Discovery

Volume 3 Issue 2
Semi Annual Edition | July - Dec, 2023
This work is licensed under CC BY-NC-SA 4.0.