Using Deep Learning and Native Mobile App to Assist Autistic
Students' Educational Experience
Zeina Thabet
a
, Hurmat Ansari
b
, Sara Albashtawi
c
, Nur Siyam
*d
and Sherief Abdallah
e
Faculty of Engineering and IT, British University in Dubai, Academic City, Dubai, U.A.E.
Keywords: Special Education, Autism, Deep Reinforcement Learning, Behaviour Intervention, Mobile Application.
Abstract: Apart from difficulties with social communication, children with autism spectrum disorder (ASD) tend to
have limited interest in academic activities. The challenges faced by the educators of these students are
abundant, including selecting motivating items or activities that can prompt them to complete a task. In
addition to this challenge, the educators also face the issue of the lack of coordination between the teachers,
therapists, and parents. This issue is imperative as significant learning opportunities are lost for lack of
communication. To address these two issues, we have created a distributed system consisting of a mobile
application that tracks the academic objectives and behavioural progress of the students which allows for a
centralized place of information for easier coordination between educators, as well as suggesting effective
motivators using a Deep Neural Network (DNN), specifically a Deep Q Network, to help autistic students
regain their focus in the class. The Deep Q Network is constructed with a custom environment that takes in
the state as input and then, based on the current state, calculates the best motivator to suggest. The mobile
application was created with an aim of assisting school educators in tracking a student’s progress. Moreover,
the system includes a staff dashboard to manage users and provide visualizations depicting students’ progress.
This project is the first of its kind and will help educators select effective motivators in moments that the
students need them as well as aid the flow of information between the stakeholders.
1 INTRODUCTION
Autism spectrum disorder (ASD) is a complex
neurodevelopmental disorder that affects 1% to 2% of
the population (CDC, 2020). ASD is distinguished by
difficulties in communication and social interactions,
repetitive and stereotyped behaviours, and restricted
interests (Schuetze et al., 2017). While there is no
known cure for ASD, early intervention has been
shown to improve cognitive abilities, language skills,
and adaptive behaviour (Dawson et al., 2012).
One difficult aspect of early intervention is
students' disinterest in academic activities or
assignments. Students with ASD may act in a
disruptive manner to avoid academic tasks (Koegel et
al., 2010). Such disruptive behaviours are regarded as
a
https://orcid.org/0000-0003-1216-9869
b
https://orcid.org/0000-0003-2725-8477
c
https://orcid.org/0000-0001-7168-6917
d
https://orcid.org/0000-0003-1625-4892
e
https://orcid.org/0000-0002-1213-2014
major impediments to the achievement of educational
goals outlined in the student's Individualized
Education Program (IEP). If untreated, disruptive
behaviours are likely to worsen.
There exists a myriad of interventions that aim to
improve core autism symptoms and academic areas.
Among those treatment with empirical support is
incorporating principles of Applied Behaviour
Analysis (ABA), which emphasizes environmental
associations and contingencies (Chasson et al., 2007;
Koegel et al., 2010; Schuetze et al., 2017).
Reinforcement learning is used in ABA-based
therapy techniques to increase desirable behaviour
(such as eye contact) and reduce atypical behaviour
(such as echoing others’ phrases) by using
motivational variables or rewards (Schuetze et al.,
96
Thabet, Z., Ansari, H., Albashtawi, S., Siyam, N. and Abdallah, S.
Using Deep Learning and Native Mobile App to Assist Autistic Students’ Educational Experience.
DOI: 10.5220/0011748800003470
In Proceedings of the 15th International Conference on Computer Supported Education (CSEDU 2023) - Volume 2, pages 96-103
ISBN: 978-989-758-641-5; ISSN: 2184-5026
Copyright
c
 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

2017). Motivators or rewards encourage children with
ASD to stay on task, follow directions, and calm
down during an outburst. Motivators can be edible
items, sensory items, activities, tokens, social
interactions, and choice. Edible items would include
things like fruits, snacks, candy, or juice. Sensory
items or activities would include activities or objects
that would simulate the senses of the student, such as
listening to music or playing with sand. Playing and
drawing are examples of possible activities. Tokens
are identified as any tangible items that are valued by
the student. Social interactions include any attention
given by another person or any interaction with
another individual, such as a teacher or parent. Choice
refers to giving the student the option to choose from
two distinct items or methods.
However, for therapists to recommend an
efficient motivator, they are required to use research-
based adaptation strategies. This includes collecting
and analysing learner’s data, update intervention
plans when there is an inadequate response,
individualized intervention based on their clinical
experience, and continuously monitor learners’
progress (Siyam, 2019). In addition, different learners
are motivated by different motivators. For example,
not all learners desire sensor stimuli nor they are all
motivated with candy (Riden et al., 2019). Thus,
identifying the right motivators for learners with ASD
is considered challenging (Mechling et al., 2006). To
solve this problem, we develop an ASD motivator
suggester powered by Deep Reinforcement Learning.
In addition to this motivator selection problem
(MSP), we also tackle the Intervention Coordination
Problem (ICP), which addresses the challenge of
sharing the information regarding the learner between
all stakeholders, including teachers, therapists, and
parents (Siyam, 2018).
The proposed solution consists of a distributed
system containing a native mobile application, a Deep
Q Network, an admin dashboard, and staff dashboard.
Through the mobile application, teachers and
therapists can track the IEP learning objectives for the
students. They can also track students’ behaviour and
use the “Motivator Selection” feature that
recommends motivators using the Deep Q Network.
Teachers can also visualize students’ progress
through the staff dashboard. Moreover,
administrators can create, edit, and delete users
records through the admin dashboard.
This work is a continuation from previous work
by Siyam and Abdallah (2021, 2022) and further
improves upon it. Previous work aimed to improve
the learning experience for students with ASD and
tackle the MSP and ICP problems using an HTML5
webapp for its functionality and a Reinforcement Q-
Learning framework for suggesting motivators.
Our implementation enhances this previous work
by migrating the functionality of motivator selection,
objective tracking, and behaviour tracking to a native
mobile application. In addition, we also upgrade the
reinforcement learning model to a deep reinforcement
learning model that has the ability to learn the best
motivator to suggest based on the current state that
has been input by the educator. Moreover, the teacher
dashboards now have visualizations that represent the
progress of students. These visualizations include
multiple types of charts and can be filtered per student
while also showing the overall achieved objectives
for all students.
To the best of our knowledge, besides our work,
there exists no other integrated system for
coordinating the effort of the teachers and therapists,
suggesting DQN predicted motivators, and progress
tracking, and visualizing for ASD affected students.
Our contributions to the previous work can be
succinctly summarized in three points:
1. Native mobile application for academic
objectives and behavioural tracking.
2. Deep Reinforcement learning AI model
(Deep Q-Network) for motivator suggestion.
3. Dashboards with visualizations that depict the
progress of students.
2 LITERATURE REVIEW
The increased prevalence of ASD diagnosis in recent
years (CDC, 2020) has fuelled machine learning
research with the goal of improving the learning
experience of those affected (Alkashri et al., 2020).
Research has primarily focused on developing
academic or social skills learning applications
(Roman et al., 2018), improving diagnosis efficiency
(Kosmicki et al., 2015), and modelling social and
behavioural aspects of ASD (Stevens et al., 2017).
Our work builds on a previous study that used
reinforcement learning to solve the problem of
selecting a motivator. In this section, we review
related work related to Deep Q Networks as well as
the use of native apps in education.
Q-Learning algorithms are model-free algorithms
that work by placing the agent within an environment
trying to find the best possible way to solve a problem
or complete a task by learning from the experience of
past actions to convert it into a policy. Q-Learning
algorithms aim to approximate the action-value
function. The Deep Q-learning algorithms do the
same thing but employ a deep neural network. The
Using Deep Learning and Native Mobile App to Assist Autistic Students’ Educational Experience
97

reason a deep neural network is needed is that when
data is highly dimensional and exists in a large state
space it is not possible to approximate an optimal
action-value function by only employing a simple q-
learning algorithm (Lazaridis et al., 2020).
Deep Q networks (DQN’s) are the primary
application of Q networks in deep learning and have
achieved superior performance compared to humans
in multiple applications. DQN’s only take the state as
input and use the neural network as a function
approximator to calculate possible outputs which are
the actions. The actions are chosen based on the
highest calculated Q-value by the neural network.
Once an action is chosen the agent performs it and
then the network is updated using new weights
calculated using the Bellman Equation (Fenjiro &
Benbrahim, 2018).
Previous studies showed that deep neural
networks (DNN’s) perform best compared to other
machine learning models due to their ability of
inferring high-level representations without the need
for extensive knowledge or preconstruction of
features. Some of the applications of DNN’s include
pre-emptive interventions, medical diagnosis, and
personalization of medication (Durstewitz et al.,
2019).
Deep Q-Learning has been used in numerous
applications, including recommendation systems in
education. For instance, Vijayan et al. (2018)
proposed a deep Q-Learning-based intelligent
learning assistant for recommending courseware to
learners with autism, based on the child's responses to
a chatbot equipped with a visual aid. The chatbot uses
multiple psychology approaches and performs an
autism assessment. The intelligent learning assistant
uses reinforcement learning and deep learning to
recommend courseware, with scores attached to each
courseware based on the child's positive or negative
response, and the deep reinforcement learning
algorithm maximizing the positive score.
3 METHODOLOGY
In this section, the methods used to implement this
work are explained.
3.1 Requirements Gathering
Design and requirements were carried over from the
previous work by Siyam and Abdallah (2021, 2022).
The authors had done extensive in-field testing and
research to attain the existing design which had
positive reviews from the ultimate users of the
system. Therefore, in this work, the website-based
app was recreated as a native app considering the
previous design so that its usability and functionality
are preserved.
3.2 Mobile Application Design
The first stage in app development was the designing
of the app. To create the prototype for the app we used
a collaborative web-based app designing tool called
Figma. We designed each page of the app using this
tool so that we would have a solid roadmap to follow
while coding the actual app. During the prototyping
process we kept the design as close to the website as
possible in terms of usability so that users of the
website who are already accustomed to its flows
would not feel disoriented when using the new app.
3.3 Creating the Deep Reinforcement
Learning Model
To implement an intelligent motivator suggester with
low human intervention, Artificial Intelligence was
utilized. Reinforcement Learning is one of three main
learning approaches in Artificial Intelligence, namely
Supervised Learning, Reinforcement Learning, and
Unsupervised Learning. A reinforcement learning
approach is an approach used when an AI agent takes
an action and then gets feedback for taking that action
which is either negative or positive feedback. An
Artificially Intelligent agent perceives its
environment and takes actions in order to reach a
specific goal. There are various types of agents in AI,
and since we are using reinforcement learning, the
type of agent we are using for this implementation is
a Learning Agent as it learns from its past experiences
in order to enhance its future performance. The
learning starts when the agent perceives a state of its
environment, and then the agent chooses an action to
perform in order to alter that current state. After that,
the agent will receive a new state of the environment
and a reward for taking the previous action that is
either positive or negative. If the reward is positive,
the agent will learn to take that same action for the
same state in the future, and if the reward is negative
the agent will learn to avoid the action for the same
state in the future (Russell & Norvig, 2016).
3.3.1 Q-Learning
The reinforcement learning algorithm used for the
code implementation is Q-Learning, where a function
Q(s,a) produces an estimate of the value of taking the
action a in the state s. We refer to this value as the Q
CSEDU 2023 - 15th International Conference on Computer Supported Education
98

value. In the beginning, the Q value will be equal to 0
as no actions were taken before for any state. When
the learning is initialized and the agent receives a state
and performs an action for that state and then receives
an award for that, two main things happen. First, the
value of Q is estimated based on the reward received
and the possible rewards that can be acquired in the
future. Second, the Q(s,a) gets updated to reflect a
new estimate after taking into account the reward
obtained using the old estimate, this allows the model
to learn from its prior experiences. The reward is an
estimation of the scores the state S receives under the
action a, which is 1 (Q-Learning Function), which is
based on Bellman’s optimality equation (see
Equation 1) (Bellman, 1966).
Q(s,a)←Q(s,a)+ α (new value estimate-Q(s,a))
(1)
The new value of Q(s,a) is the sum of the old
Q(s,a) and an updating value. This value is calculated
by multiplying the difference between the old and
new values by a learning coefficient α. When α=0,
there will be no updated value, and when α=1, the
new value is taken for Q(s,a) while the old value gets
ignored entirely. By adjusting the value of α, the
speed of updating previous knowledge with new
knowledge is determined. Q-values are stored in a Q-
table as a history record so that it can be used for
future states.
3.3.2 Exploration vs Exploitation
Since deep reinforcement learning models do not
have datasets to learn from, they must create their
own data to learn. This learning happens when the AI
agent explores its environment to learn all the
possibilities of actions to take in order to maximize
reward. Exploitation is when the agent knows to
apply its learning from exploration in order to take
previously experienced actions to maximize rewards.
3.3.3 Deep Q-Learning
Although Q-learning is a powerful algorithm on its
own, it does suffer some limitations due to the fact
that this method is slow and is restricted to previous
experiments. Despite adding exploration to it, the Q
function lacks flexibility. This was the reason we
decided on developing a Deep Q-Learning algorithm.
A deep Q-Learning algorithm will make up for the
shortfalls of a standard Q-Learning algorithm. After
training, Deep Learning algorithms would be to take
the best action to a state it has never seen before,
which is considered better than classic Q-Learning
that is only limited to a set list of states.
Deep Q-Learning utilizes neural networks as
opposed to a Q-table. The state is fed into the neural
network which calculates the q-value corresponding
to each action. The best action is the one
corresponding to the highest q value. The input layer
in the neural network is the same size as the states and
the output layer is the same size as the possible
actions that can be taken. The inputs are inserted in
the neural network which then outputs q-values that
correspond to each action taken for that state. Once
the state is input an episode begins and it continues
until the agent reaches the terminal state which is
when the reward related to the action taken is
acquired. Episodes do not affect each other. However,
the agent does learn from each episode which in turn
makes the agent choose better actions with higher
rewards in subsequent episodes.
3.3.4 Stable Baselines 3
The library we have used to implement the Deep Q
Network is Stable Baselines 3 which is a library used
for reinforcement learning implementation.
3.3.5 Custom Gym Environment
In order to build the Deep Q-Learning algorithm, we
have to create an environment. The environment is
the world which contains the observation space and
where the actions happen. In our case the
environment consists of observation space, action
space, and reward. The observation space is the list of
possible states which in our case are Trigger, Time of
Day, Subject, Behaviour, and Behaviour Function.
Trigger refers to the reason that caused a student to
behave in a certain way. Time of Day refers to the
time the student displayed a behaviour during the day.
Behaviour is the conduct of the student, and
Behaviour function refers to the desired target
intended to be obtained by exhibiting a specific
behaviour. The action space is the list of possible
actions that can be taken for any state. Then the
reward is calculated after each action is taken. There
is a criterion for calculating the reward, which is
based on the change in state. The better the change in
state, the higher the reward and vice versa (Siyam &
Abdallah, 2022).
When the teacher inputs a state and asks for a
recommended action, the system sends a
recommendation. If the teacher declines the
recommendation, nothing happens and the reward is
discounted, but if they accept the recommendation,
they give the student the recommended motivator.
After that, the system waits for feedback which is
then used to calculate the reward. The criterion for
Using Deep Learning and Native Mobile App to Assist Autistic Students’ Educational Experience
99

Figure 1: Communication process between the DQN model, the server, and the mobile application.
calculating the reward follows the research by Siyam
and Abdallah (2022).
The definition of safe Reinforcement Learning
was previously introduced in order to balance the
success of the motivator suggester with the long-term
avoidance of potentially harmful recommendations
like sugary treats and violent movies (Siyam &
Abdallah, 2022).
3.4 Creating the Server
Django framework was used to create a server that
stores the data and to be able to send it over to the
client (mobile application). The server follows the
REST architecture where requests are sent to the
server and responses are sent back to the client.
Several different types of requests can be received
over to the server, and they include GET, POST, and
PATCH requests.
The server includes many different endpoints to
serve data needed to the client and also includes
endpoints that send suggested motivators from the
Deep Q Networks model to the client.
3.4.1 DQN Structure Server-side
The communication method between the server and
the mobile application are HTTP requests and
responses, whereas the communication method
between the server and the DQN model are socket
connections. The whole process is a two-stage
process, where the first stage starts by the user
requesting a motivator suggestion by providing
information about the state. This data gets sent over
to the server and is then sent to the DQN model. The
DQN model receives the state data as input and
outputs a motivator suggestion which then gets sent
back and gets saved in the database along with the
provided state data. The client receives the suggested
action and is displayed to the user (see Figure 1).
The second stage starts when the user sends
feedback about the suggested motivator to the server,
where the server then retrieves the previously saved
state data using the motivator ID and all that
information gets sent over to the DQN model to allow
it to learn by calculating the reward according to the
feedback it received. The results are then stored in the
database.
3.5 Building the App
The native mobile app is built using an open-source
mobile user interface framework called Flutter.
Flutter provides the ability to create cross platform
apps using one codebase, which is why it was chosen
as the implementation framework. Once the app was
created it was manually tested by group members to
ensure its usability.
The mobile application consists of different
screens to allow the user to view their students, view
their student’s objectives, and add, edit, delete, or
update objective updates. As well as add behaviour
updates and the addition of notes under objective
updates (see Figure 2).
3.6 Building the Dashboards
Alongside the mobile application, two dashboards
were created which consisted of a staff dashboard and
an admin dashboard. The staff dashboard displays
CSEDU 2023 - 15th International Conference on Computer Supported Education
100

Figure 2: Mobile app screenshots.
visualizations related to the student progress to allow
the staff member to view the overall progress of the
student in an easy to comprehend manner. Whereas
the admin dashboard allows the admin user to add,
edit, update, and delete data related to the different
users, staff, parents, and students. The Chart.js library
was used to construct such visualizations from the
student data stored in the database.
Figure 3: Teacher's dashboards.
3.6.1 Staff Dashboard
The staff dashboard included various chart that allow
teachers and therapists to view the progress of different
objectives across different types of objectives,
percentage of objectives achieved, and the number of
achieved objectives per student (see Figure 3).
The dashboard includes several graphs which
include line graphs, a bar chart and a tree map.
The line graph that measures the progress of the
student’s objectives across time. This helps in getting
an overview of the students’ progress and whether
they are progressing well or not.
The bar chart lists the number of completed
objectives vs non completed objectives, where this
allows the teacher to gauge how close they are to
completing their objectives.
And the tree map illustrates the number of
completed objectives per student, where this can help
the teacher identify high performing students and
students who are lacking behind in achieving their
objectives.
3.6.2 Admin Dashboard
The admin dashboard allows full control over the data
stored in the database as well as profile editing. This
allows administrators to have authority to execute
administrative tasks in the application.
4 EVALUATION
The evaluation of the proposed system depends on its
usage by educators at schools with learners with
ASD. In previous work, the mobile app has been
developed and tested as an app that facilitates
communication and coordination between different
parties involved in the therapy and learning of
students with ASD (Siyam, 2021; Siyam & Abdallah,
2022). In this paper, we describe the evaluation
process for the proposed deep reinforcement
algorithm. In comparison to other machine learning
algorithms, reinforcement lacks an agreed-upon
performance evaluation standard (Liu et al., 2020).
While this is not a problem unique to RL, it is more
difficult to address when compared to other machine
learning algorithms that use accuracy and precision
recall as performance indicators. To calculate an
algorithm's precision and accuracy, an offline dataset
must be divided into training and testing sets. Because
this study lacks an offline dataset, we propose
evaluating the effectiveness of the algorithm through
statistical and qualitative methods (Stratton, 2019).
Therefore, the average reward per episode will be
used to gauge the improvement in performance of the
proposed DQN. With every passing episode, the
average reward should be increasing to show that the
DQN is learning to take better actions that beget
better rewards.
Siyam and Abdallah (2022) evaluated their Q-
Learning model by comparing the effectiveness of the
motivators when suggested by the model vs random
suggestions. And concluded that the Q-Learning
model was beneficial due to much higher
effectiveness recorded from the users of the model.
Building up on that research, we will be recording the
DQN’s effectiveness and comparing it to the Q-
learning effectiveness, where if our DQN model
illustrates higher effectiveness than the Q-learning
model by Siyam and Abdallah (2022), our work will
Using Deep Learning and Native Mobile App to Assist Autistic Students’ Educational Experience
101

be proven to be an enhancement over the previous
iteration.
5 CONCLUSION
In this work, we have developed a native app which
would offer more accessibility compared to the
previous web-based application. Moreover, we
employed deep reinforcement learning using neural
networks to improve the recommendations even
when new scenarios arise.
The primary goal of this work is to aid the
education of learners with ASD. This goal was
achieved by the creation of a distributed system
consisting of a native app that tracks the objective and
behavioural progress of the students, as well as,
suggesting motivators using a Deep Q-network.
Moreover, the system includes dashboards for
administrators and teachers at the educational
institution. The admin dashboard provides the
functionality for editing, adding, or deleting users,
while the teacher’s dashboard allows them to view
their students’ progress by presenting data
visualizations that illustrate it. The server allows to
have a way of communication between the database
and the mobile app.
The native app for this system was designed using
Figma, which is a prototyping graphic designing tool,
and actualized using an open-source mobile user
interface framework called Flutter. Flutter allows the
creation of a cross platform apps using one codebase
which is why it was chosen as the implementation
framework. This will also allow a higher number of
users to be able to use the native app rather than the
web-based app.
The Deep reinforcement learning motivator
suggester was implemented using the Python library
Stable Baselines3 (STB3) in addition to a custom
environment that we created using the library gym. A
custom environment was needed and constructed to
identify observation space, action space, and reward
while the STB3 library was used to construct the
Deep Q-network.
6 LIMITATIONS
This work has some limitations that should be
considered. For instance, the proposed algorithm does
not suggest motivators customized to each student.
The best action suggestion made by the DQN
algorithm only takes into consideration the current
state with the criteria Trigger, Time of Day, Subject,
Behaviour and Behaviour Function. This treats the
student body as a monolithic entity and suggests
motivators based on impersonalized factors. Another
limitation of this work is that notifications are not sent
to the stakeholders when objectives are achieved or
when there is progress in behavioural goals.
7 FUTURE WORK
Further improvements can be made in the project on
mobile app experience and features level as well as
on the deep learning model level.
The mobile application can be improved by
adding notifications feature that will allow the
stakeholders to be updated in a timely order on
student’s objectives progress. These notifications will
notify users about the most pertinent updates as
opposed to every single update. Moreover, it could be
improved by adding more visualizations such as a
double bar chart where each double bar represents a
teacher that depicts the number of completed and
non-completed objectives for that teacher’s students.
In addition, adding a reminder system to remind
different stakeholders of students who have not had
much progress in their objectives, identified using
several data analysis methods, could prove to be
beneficial in helping students quickly identify them
and direct focus to them.
On the other hand, the Deep Q Network can be
improved by recommending motivators to be
customized per specific learner instead of a
generalized recommendation.
Finally, the system should be deployed in a school
environment so that it can be used by teachers,
therapists, and staff at a school. Once deployed, in-
field testing can be conducted to allow the Deep Q-
Network to learn from continual use so that its
recommendations can become more effective. The
results acquired from the testing will then be
evaluated using both DQN performance evaluation,
and results and reviews from the stakeholders using
the system at the institution. The review, feedback,
and recommendations compiled from the
stakeholders will be considered to further enhance,
strengthen, and streamline the system and its user
experience.
REFERENCES
Alkashri, Z., Siyam, N., & Alqaryouti, O. (2020). A
detailed survey of Artificial Intelligence and Software
CSEDU 2023 - 15th International Conference on Computer Supported Education
102

Engineering: Emergent Issues. 2020 Fourth
International Conference on Inventive Systems and
Control (ICISC), 666–672. https://doi.org/10.1109/
ICISC47916.2020.9171118
Bellman, R. (1966). Dynamic Programming.
Science,153(3731), 34–37. https://doi.org/10.1126/
science.153.3731.34
CDC. (2020). Data and Statistics on Autism Spectrum
Disorder | CDC. Centers for Disease Control and
Prevention. https://www.cdc.gov/ncbddd/autism/
data.html
Chasson, G. S., Harris, G. E., & Neely, W. J. (2007). Cost
Comparison of Early Intensive Behavioral Intervention
and Special Education for Children with Autism.
Journal of Child and Family Studies, 16(3), 401–413.
https://doi.org/10.1007/s10826-006-9094-1
Dawson, G., Jones, E. J. H., Merkle, K., Venema, K., Lowy,
R., Faja, S., Kamara, D., Murias, M., Greenson, J.,
Winter, J., Smith, M., Rogers, S. J., & Webb, S. J.
(2012). Early Behavioral Intervention Is Associated
With Normalized Brain Activity in Young Children
With Autism. Journal of the American Academy of
Child & Adolescent Psychiatry, 51(11), 1150–1159.
https://doi.org/10.1016/j.jaac.2012.08.018
Durstewitz, D., Koppe, G., & Meyer-Lindenberg, A.
(2019). Deep neural networks in psychiatry. Molecular
Psychiatry, 24(11), Article 11. https://doi.org/10.
1038/s41380-019-0365-9
Fenjiro, Y., & Benbrahim, H. (2018). Deep reinforcement
learning overview of the state of the art. Journal of
Automation Mobile Robotics and Intelligent Systems,
Vol. 12, No. 3. https://doi.org/10.14313/JAMRIS_3-
2018/15
Koegel, L., Singh, A. K., Koegel, R. L., & Koegel, L.
(2010). Improving Motivation for Academics in
Children with Autism. Journal of Autism and
Developmental Disorders, 40(9), 1057–1066.
https://doi.org/10.1007/s10803-010-0962-6
Kosmicki, J. A., Sochat, V., Duda, M., & Wall, D. P.
(2015). Searching for a minimal set of behaviors for
autism detection through feature selection-based
machine learning. Translational Psychiatry, 5(2), e514.
https://doi.org/10.1038/tp.2015.7
Lazaridis, A., Fachantidis, A., & Vlahavas, I. (2020). Deep
Reinforcement Learning: A State-of-the-Art
Walkthrough. Journal of Artificial Intelligence
Research, 69, 1421–1471. https://doi.org/10.1613/
jair.1.12412
Liu, S., See, K. C., Ngiam, K. Y., Celi, L. A., Sun, X., &
Feng, M. (2020). Reinforcement Learning for Clinical
Decision Support in Critical Care: Comprehensive
Review. Journal of Medical Internet Research, 22(7),
e18477. https://doi.org/10.2196/18477
Mechling, L. C., Gast, D. L., & Cronin, B. A. (2006). The
Effects of Presenting High-Preference Items, Paired
With Choice, Via Computer-Based Video
Programming on Task Completion of Students With
Autism. Focus on Autism and Other Developmental
Disabilities
, 21(1), 7–13. https://doi.org/10.1177/
10883576060210010201
Riden, B. S., Markelz, A. M., & Randolph, K. M. (2019).
Creating Positive Classroom Environments With
Electronic Behavior Management Programs. Journal of
Special Education Technology, 34(2), 133–141.
https://doi.org/10.1177/0162643418801815
Roman, J., Mehta, D. R., & Sajja, P. S. (2018). Multi-agent
Simulation Model for Sequence Generation for
Specially Abled Learners. In S. C. Satapathy & A. Joshi
(Eds.), Information and Communication Technology
for Intelligent Systems (ICTIS 2017)—Volume 1 (pp.
575–580). Springer International Publishing.
Russell, S., & Norvig, P. (2016). Artificial Intelligence: A
Modern Approach, Global Edition. Pearson.
Schuetze, M., Rohr, C. S., Dewey, D., McCrimmon, A., &
Bray, S. (2017). Reinforcement Learning in Autism
Spectrum Disorder. Frontiers in Psychology, 8.
https://doi.org/10.3389/fpsyg.2017.02035
Siyam, N. (2018). Special Education Teachers’
Perceptions on Using Technology for Communication
Practices. https://bspace.buid.ac.ae/handle/1234/1349
Siyam, N. (2019). Factors impacting special education
teachers’ acceptance and actual use of technology.
Education and Information Technologies, 24(3), 2035–
2057. https://doi.org/10.1007/s10639-018-09859-y
Siyam, N. (2021). Using Mobile Technology for
Coordinating Educational Plans and Supporting
Decision Making Through Reinforcement Learning in
Inclusive Settings [Thesis, The British University in
Dubai (BUiD)]. https://bspace.buid.ac.ae/handle/
1234/1879
Siyam, N., & Abdallah, S. (2021). A Pilot Study.
Investigating the Use of Mobile Technology for
Coordinating Educational Plans in Inclusive Settings.
Journal of Special Education Technology, 1–14.
https://doi.org/10.1177/01626434211033581
Siyam, N., & Abdallah, S. (2022). Toward automatic
motivator selection for autism behavior intervention
therapy. Universal Access in the Information Society.
https://doi.org/10.1007/s10209-022-00914-7
Stevens, E., Atchison, A., Stevens, L., Hong, E.,
Granpeesheh, D., Dixon, D., & Linstead, E. (2017). A
Cluster Analysis of Challenging Behaviors in Autism
Spectrum Disorder. 2017 16th IEEE International
Conference on Machine Learning and Applications
(ICMLA), 661–666. https://doi.org/10.1109/ICMLA.
2017.00-85
Stratton, S. J. (2019). Quasi-Experimental Design (Pre-Test
and Post-Test Studies) in Prehospital and Disaster
Research. Prehospital and Disaster Medicine.
Cambridge University Press, vol. 34(6), pp. 573–574
Vijayan, A., Janmasree, S., Keerthana, C., & Baby Syla
(2018). A Framework for Intelligent Learning Assistant
Platform Based on Cognitive Computing for Children
with Autism Spectrum Disorder. 2018 International
CET Conference on Control, Communication, and
Computing (IC4), 361–365. https://doi.org/10.1109/
CETIC4.2018.8530940.
Using Deep Learning and Native Mobile App to Assist Autistic Students’ Educational Experience
103