AI Tutor: Adaptive e-Learning System Using Expert Fuzzy

Controllers

Marcin Szczepański

, Grzegorz Gapiński and Jacek Marciniak

Faculty of Mathematics and Computer Science, Adam Mickiewicz University Poznań,

Uniwersytetu Poznańskiego 4 Street, Poznań, Poland

Keywords: e-Learning Course with Adaptive Content, Adaptive e-Learning System, Expert Fuzzy Controller.

Abstract: AI Tutor is an e-learning system that adapts content to each student’s unique learning style. Achieving this

level of adaptability requires specialized methods, and the solution presented here employs concepts

of information imprecision and fuzzy expert control. Within an e-learning course titled “Introduction

to Machine Learning” in an Artificial Intelligence curriculum, three fuzzy controllers were specifically

designed and implemented to adjust learning materials in real-time. This personalized approach highlights

the strength of fuzzy controllers in e-learning, allowing the course to effectively respond to a wide range of

learning preferences. By addressing the imprecision in how information is processed and understood,

these controllers handle the variability and uncertainty inherent in individual learning styles. Ultimately,

AI Tutor demonstrates the potential of fuzzy logic to enhance adaptive e-learning, creating a more tailored

and effective learning experience for students with diverse needs.

1 INTRODUCTION

The development of personalized e-learning systems

is one of the key challenges in modern education. In

an era of rapidly changing student needs and diverse

learning styles, adaptive technologies that enable the

tailoring of educational content to individual user

requirements are playing an increasingly important

role. Building such systems requires flexible

approaches that take into account the complexity of

educational processes and the imprecision of data on

student behavior. These systems can be developed

using different techniques, such as rule-based systems

that leverage expert knowledge, or machine learning

methods that learn from large data sets (Caro et al.,

2015; Fenza et al., 2017). Among these solutions, an

encouraging alternative is advanced fuzzy controllers

capable of dynamically adapting educational content

in response to individual student interactions.

Fuzzy controllers based on expert knowledge

provide an alternative to traditional machine learning

methods, which typically require large data sets for

training. Unlike these methods, fuzzy controllers rely

on rules developed from teachers' experience and

designed specifically to incorporate imprecise

https://orcid.org/0000-0002-6185-6115

https://orcid.org/0000-0002-1186-9612

information. The ability to process such information

is critical in the educational process because many

phenomena that teachers consider are based on

numerous factors that are difficult to define precisely,

such as student motivation, the pace of material

assimilation, or individual learning challenges

(Kasinathan et al., 2017; Santos et al., 2020). By

using fuzzy modeling, adaptive e-learning systems

can effectively account for these complex and

subjective aspects when personalizing learning paths.

The aim of this article is to present the concept

and application of fuzzy controllers in adaptive e-

learning systems. Special emphasis is placed on

analyzing their ability to adapt content based on

variable and diverse student behavior. A solution is

presented in which three different fuzzy controllers

were implemented within a single e-learning course

to enable adaptation that takes into account student

progress and engagement. Depending on the

personal learning advancement, the course takes

different forms and adapts to the identified needs.

An important feature of the proposed solution is the

ease of interpretation of its behavior by the teachers,

thanks to the transparency of the rules constructed,

which makes it easier to adjust the behavior of the

Szczepa

nski, M., Gapi

nski, G. and Marciniak, J.

AI Tutor: Adaptive e-Learning System Using Expert Fuzzy Controllers.

DOI: 10.5220/0013277900003932

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 17th International Conference on Computer Supported Education (CSEDU 2025) - Volume 2, pages 96-107

ISBN: 978-989-758-746-7; ISSN: 2184-5026

system whenever the teacher identifies a need for

modification.

2 BACKGROUND

Research on adaptive content and adaptive learning

systems has been conducted since computers were

first introduced in education (Böcker et al., 1990).

Beyond educational settings, adaptive content is also

valuable in areas such as marketing, e-commerce, and

recommendation systems (Casillo et al., 2021; Desai,

2022; Vinaykarthik and Mohana, 2022). The

development of adaptive content is critical for

personalizing the learning experience, which is

essential in modern education.

E-learning courses with adaptive content are

defined by their ability to adjust material based on

factors like the learner’s individual preferences or

progress within the course (Dorça et al., 2017;

Ennouamani and Mahani, 2017; Premlatha and

Geetha, 2015; del Puerto Paule Ruiz et al., 2008).

Several solutions leveraging learning styles for

content adaptation in adaptive e-learning courses are

found in the literature. For instance, in 2017, Fabiano

A. Dorça and colleagues developed a solution

recommending additional content based on a pre-

defined ontology that links relationships among

learning objects to learning styles in the Felder-

Silverman model (Dorça et al., 2017). Similarly, in

2019, Nisha S. Raj and Renumol V. G. proposed an

adaptation approach grounded in the Felder-

Silverman model, delivering course content

according to a rule-based system (Raj and V G, 2019).

In 2021, Hassan A. El-Sabagh introduced a method

for identifying learning styles using the VARK model

(Fleming, 2006; El-Sabagh, 2021). His study also

demonstrated that adapting content based on learning

styles had a statistically significant positive impact on

student engagement, measured by Marcia Dixson’s

48-item engagement scale, assessing skills,

interaction, performance, and emotional engagement

(Dixson, 2015).

Adaptation algorithms are not limited to learning

styles alone. For instance, in 2017, Giuseppe Fenza,

Francesco Orciuoli, and Demetrios Sampson (Fenza

et al., 2017) proposed a solution that uses a neural

network model trained on data, including inputs from

educators. This model generates rules that shape the

format of the next task for the student. These rules are

derived from the student’s actions in previous tasks.

Fuzzy logic, including fuzzy control methods, is

also applied in designing e-learning courses with

adaptive content (Chandrasekhar and Khare, 2021;

Marciniak et al., 2023; Szczepański and Marciniak,

2023). Fuzzy controllers rely on expert knowledge,

meaning that the rule base is always developed by

specialists in the specific field where the system has

some imprecise problems to solve (Zadeh, 1965).

Fuzzy controllers are applied in situations where

decisions must be made despite incomplete data or

when creating too many rules in a rule-based system

is impractical (Mendel, 2017). Their flexibility allows

them to manage imprecise or uncertain data,

representing it as degrees of membership rather than

binary values (Khomeiny et al., 2020). This makes

them ideal for adaptive learning systems and other

applications, where input data is often unclear or

uncertain (Kovacic and Bogdan, 2018). Fuzzy

controllers can be used as independent applications or

integrated into a more comprehensive adaptive

learning system.

The adoption of adaptive content and adaptive

learning systems in education has been increasing in

recent years. These systems have proven effective in

meeting the needs of diverse learners and offering

personalized learning experiences (Katsaris and

Vidakis, 2021). As digital technologies continue to be

widely implemented in education, the demand for

adaptive learning systems and content is expected to

rise in the future (Sushama et al., 2022). A major

advantage of adaptive learning systems is their ability

to provide real-time feedback and personalized

support, allowing learners to advance at their own

pace (Lerís et al., 2017). These systems also offer

instructors valuable data on learners’ progress,

helping them pinpoint areas where extra assistance

may be required. This information enables instructors

to adjust their teaching methods to better meet the

individual needs of learners and enhance the overall

learning experience (Gaudioso et al., 2012).

3 AI TUTOR COURSE WITH

ADAPTIVE CONTENT

The AI Tutor represents an example of an adaptive e-

learning system that dynamically adapts the content

of an e-learning course according to a pre-defined

adaptation strategy. This strategy has been

implemented in the course “Introduction to Machine

Learning”, which is part of the instructional toolkit

used in an Artificial Intelligence curriculum.

Designed to introduce the fundamentals of machine

learning through practical examples and exercises,

this course engaged 89 computer science students

enrolled in the course.

AI Tutor: Adaptive e-Learning System Using Expert Fuzzy Controllers

3.1 Course Structure

Aligned with the Universal Curricular Taxonomy

System (UCTS) (Marciniak, 2014), the “Introduction

to Machine Learning” course is structured as a single

UCTS Module. This module comprises four UCTS

Units, each containing three to eight Learning

Objects, including at least one dedicated to review,

and is followed by a skills-oriented assessment. At the

beginning of the course, students complete a

diagnostic test to evaluate their theoretical knowledge

of machine learning. At the course’s conclusion, they

may take a final skills-oriented test to potentially

improve their scores from prior assessments. A

screenshot of the initial knowledge-oriented test is

shown in Figure 1, while Figure 2 shows an example

fragment of the skills-oriented assessment at the end

of the “Metrics” UCTS Unit.

Figure 1: Initial (knowledge-oriented) test in the

"Introduction to Machine Learning" course.

Figure 2: Fragment of the end-of-unit skills-oriented

assessment in the "Introduction to Machine Learning"

course.

"Introduction to Machine Learning” was

developed using the Eduexe e-learning authoring tool

(Eduexe, 2024), with references to Google Teachable

Machine (Google Teachable Machine, 2019) used to

design practical examples and exercises. The course

was produced as a package in the SCORM standard

with Eduexe platform extensions, allowing the

implementation of a fuzzy rule-based system. It was

made available to students through Moodle. The

course was self-paced and students had six days to

complete it. It contained 19 substantive learning

objects as well as tests, questionnaires, and technical

learning objects that serve an informational function

and facilitate navigation through the course. A

detailed structure of the course, following the UCTS

framework, is shown in Table 1.

Table 1: Structure of “Introduction to Machine Learning”

course in UCTS framework.

Course part No. of

Learning

objects

UCTS

taxon

Introduction to Machine

Learnin

Module

Initial diagnostic test

(

knowled

e-oriente

)

Exam

Introduction 3 Unit

Data in the process of

learning

3 Unit

Basic concepts of

Machine Learnin

5 Unit

Metrics 8 Unit

Final test (skills-

oriente

)

Exam

An example, an intentionally imperfect Machine

Learning model used in the course, prepared using the

Google Teachable Machine tool, is shown in Figure 3.

Figure 3: Example Machine Learning model used in the

course.

3.2 Content Adaptation Strategy

The AI Tutor system employs a hybrid adaptive

strategy, leveraging various content adaptation

techniques to achieve three main objectives: (1)

maintaining engagement among high-performing

students, (2) supporting lower-performing students

with additional review materials and exercises, and

CSEDU 2025 - 17th International Conference on Computer Supported Education

(3) minimizing situations where high final scores may

not accurately reflect a student's competence level.

The first objective is achieved by continuously

measuring each student's Machine Learning

Competence indicator, defined as a combination of

theoretical knowledge and practical skills. This

measurement process includes an initial knowledge-

oriented test and skills-oriented assessments at the

end of each UCTS Unit. A fuzzy controller

synthesizes these results into a Machine Learning

Competence score standardized on a scale from zero

to one. If a student’s score reaches or exceeds 0.8,

they are granted unrestricted access to course content,

allowing them to complete units in any sequence

without needing to complete review tasks or end-of-

unit assessments in the order imposed by the course.

This adaptation minimizes repetitive tasks for

advanced students, aiming to sustain their

engagement by preventing boredom.

The adaptive strategy also addresses the needs of

students who struggle. If a student with moderate or

low Machine Learning Competence fails an end-of-

unit assessment, they are directed to supplementary

examples and review tasks. Completing these tasks is

mandatory before retaking the assessment and

progressing in the course.

The third goal of the adaptive strategy seeks to

reduce the possibility of a student with low Machine

Learning Competence scoring unexpectedly high on

the final assessment. Such results could indicate a

possible compromise of the test question pool, which

may occur in the case of an educational process

involving a course that lasts several days and is

conducted as self-study without teacher supervision.

In a simpler content adaptation strategy implemented

in another previously developed course, the use of the

disengagement phenomenon was proposed to deliver

question pools of the same difficulty level in a

controlled manner to users with different levels of

learning disengagement (Szczepański and Marciniak,

2023). Disengagement is standardized on a scale from

zero to one and is calculated using a fuzzy logic

controller that processes inputs related to the quality

of student learning (influenced by the frequency of

interactions with course elements and the time spent

on individual learning objects) and the time

remaining before course access concludes

(Szczepański and Marciniak, 2023). Low

disengagement scores indicate sustained student

effort and consistent engagement with the course

content, while high disengagement scores suggest a

lack of engagement, with students either neglecting

course tasks or engaging with them superficially,

often when course deadlines approach.

However, it is also possible that a high

disengagement indicator is the result of too low a

level of course difficulty relative to the high level of

user competence – in such a case, this indicator alone

should not be the sole premise for the decision to

replace the pool of test questions. To ensure a fair and

accurate assessment, the AI Tutor system employs a

Question Exchange Requirement (QER) indicator,

which assigns a value between zero and one based on

fuzzy controller outputs from both the Machine

Learning Competence and disengagement

controllers. If a student’s QER indicator exceeds 0.5,

the final test questions are selected from an

alternative pool of questions of equivalent difficulty

but varied content, providing a robust and reliable

measure of each student's mastery. Otherwise,

questions in the final test are drawn from the pools of

questions from the assessments summarizing each

UCTS Unit – some questions may be repeated in this

way. This is a form of bonus for committed students,

helping them improve their final score in the course –

however, it is not a discretionary bonus, because it is

justified by the Machine Learning Competence

developed during the course and by the commitment

to perform additional and revision tasks.

4 ADAPTIVE E-LEARNING

SYSTEM WITH EXPERT

FUZZY CONTROLLERS

The instructional phenomena used in the AI Tutor

system’s content adaptation framework are

characterized by imprecision. To address this, the

severity of each phenomenon is categorized into

levels defined as low, medium, or high, with these

distinctions based on various input parameters. The

determination of these severity levels is based on a set

of rules defined by domain experts i.e. experienced

teachers. These experts construct a structured rule

base, which takes the form of if…then conditional

statements that articulate the relationship between

specific input variables and the respective

instructional phenomena being modeled.

This rule-based framework serves as the basis for

aligning system behavior with expert knowledge.

Given the intrinsic ambiguity and variability of the

instructional concepts under consideration, as well as

the expert-driven nature of the rule base, Mamdani's

fuzzy inference model (Mamdani, 1974) was selected

as the most appropriate approach to accurately

capture and model these instructional phenomena.

Mamdani's fuzzy controller facilitates a nuanced

AI Tutor: Adaptive e-Learning System Using Expert Fuzzy Controllers

representation of imprecise relationships by applying

fuzzy logic principles, allowing the AI Tutor system

to emulate expert decision-making in content

adaptation with a high degree of flexibility and

interpretability.

4.1 Machine Learning Competence

According to the content adaptation strategy detailed

in Section 3.2, this approach incorporates an

evaluation of the student's competence in

fundamental machine learning concepts (Machine

Learning Competence). This competence level is

estimated using Mamdani’s expert fuzzy controller,

which processes two input variables: (1) the student’s

baseline knowledge level, represented by their score

on an initial diagnostic test (knowledge), and (2) their

skill level in machine learning fundamentals,

reflected in scores obtained from assessments

following each course unit (skill).

For each input variable, three linguistic values

(terms) were defined: low, medium, and high.

Minimizing the number of terms and variables

simplifies the rule base, making it more accessible

and interpretable for the expert – in this case, the

teachers. Figure 4 illustrates the membership

functions associated with the fuzzy sets representing

these linguistic terms.

Figure 4: Model of variables values in the controller

assessing Machine Learning Competence.

Consistent with the Mamdani fuzzy controller

model, an output variable was established within the

defuzzification module to represent the student’s

overall competence level in machine learning

fundamentals (competence). This output variable was

also defined using three terms, mirroring the structure

of the input variables (refer to Figure 4).

The definitions of these terms align with the

course’s grading criteria: a student is deemed to have

failed if they score below 50% across course

activities, while a score of 75% or above indicates

high performance. This alignment ensures that the

model’s output reflects real-world academic

evaluations.

The subsequent step involved developing a rule

base, as shown in Table 2, which lists all the fuzzy

rules applied in the controller. When constructing the

rule base, it was assumed that, within the context of

machine learning fundamentals, skills are prioritized

over knowledge. For instance, in Rule 3, if the

knowledge level is low, but the skill level is high, the

overall competence is rated as medium. However, in

the reverse situation (Rule 7), where the knowledge

level is high, but the skill level is low, the competence

remains low.

Table 2: Fuzzy controller rule base for calculating Machine

Learning Competence.

No. Rule

knowledge is low and skill is low then

competence is low

knowledge is low and skill is medium then

competence is low

knowledge is low and skill is high then

competence is mediu

knowledge is medium and skill is low then

competence is low

knowledge is medium and skill is medium

then competence is mediu

knowledge is medium and skill is high then

competence is hi

knowledge is high and skill is low then

competence is low

knowledge is high and skill is medium then

competence is mediu

knowledge is high and skill is high then

competence is hi

In the rule antecedents, the two input variables are

connected by a logical conjunction (and), modeled as

a minimum operation in the controller, in line with

common fuzzy logic practices. The implication (then)

operator is also defined as a minimum operation.

Through the fuzzy inference process based on this

rule base, a composite fuzzy set is generated by

summing the individual fuzzy sets produced by each

rule for specific input values. The final step in the

fuzzy controller’s process involves defuzzifying this

composite set, with the center of gravity method used

to yield a precise output – a widely preferred method

for defuzzification in fuzzy systems.

CSEDU 2025 - 17th International Conference on Computer Supported Education

100

4.2 Disengagement

A further instructional phenomenon incorporated into

content adaptation strategies is student

disengagement. This phenomenon is also modeled

using Mamdani’s expert fuzzy controller, which

evaluates two input variables: the student’s learning

quality within the course (learning_quality) and the

time remaining until the final test, measured from the

point at which the student began the course

(remaining_time). This controller has previously

been successfully implemented in another e-learning

course which employed a simpler content adaptation

strategy (Szczepański and Marciniak, 2023).

The first input variable, the student’s learning

quality, is defined as an aggregate measure based on

three indicators of the student’s progress in the course

(Szczepański and Marciniak, 2023): the number of

interactive exercises completed within the learning

materials (interactions), the average time spent on

each segment of content (learning object) in a given

unit (time), and the number of content segments

(learning objects) in the unit that were not accessed

by the student (not_visited). The calculation method

for learning quality is provided in Equation (1).

learning_quality =

interactions + 2 ∙ time - not_visited

(1)

A teacher with in-depth knowledge of their

students often encounters difficulties in objectively

quantifying the factors that influence learning quality.

To address this challenge, it is essential to establish a

method for the teacher to quantitatively evaluate

various aspects of student behavior. The solution

proposed here entails creating a formula refined

through iterative analysis of student data from prior

courses that did not utilize a fuzzy controller. This

approach aggregates relevant data, enabling the use of

two linguistic variables, which significantly

simplifies the rule base construction, requiring only

nine rules. Without this data aggregation, the system

would need to manage four input variables,

potentially expanding the rule base to as many as 81

rules (Szczepański and Marciniak, 2023).

The second input variable for the fuzzy controller,

which calculates the phenomenon of disengagement,

is the normalized time remaining until the final test

deadline, measured from the point at which the

student begins engaging with the course.

Additionally, the controller defines an output variable

– student’s disengagement (disengagement). Each

variable is expressed through three linguistic values:

low, medium, and high. The membership functions

that map these values to their corresponding fuzzy

sets are shown in Figure 5.

Figure 5: Model of variables values in the controller

assessing student’s disengagement.

Table 3: Fuzzy controller rule base for calculating student’s

disengagement (Szczepański and Marciniak, 2023).

No. Rule

if learning_quality is low and

remaining_time is low then disengagement

is hi

if learning_quality is low and

remaining_time is medium then

disen

ement is mediu

if learning_quality is low and

remaining_time is high then disengagement

is mediu

if learning_quality is medium and

remaining_time is low then disengagement

is mediu

if learning_quality is medium and

remaining_time is medium then

disen

ement is mediu

if learning_quality is medium and

remaining_time is high then disengagement

is low

if learning_quality is high and

remaining_time is low then disengagement

is mediu

if learning_quality is high and

remaining_time is medium then

disen

ement is low

if learning_quality is high and

remaining_time is high then disengagement

is low

The rule base of the fuzzy controller consists of

nine rules, as outlined in Table 3. Similar to the fuzzy

controller described in Section 4.1, the conjunction

operator (and) is implemented using the minimum

AI Tutor: Adaptive e-Learning System Using Expert Fuzzy Controllers

101

operation, as is the implication operator (then). The

fuzzy set produced by the controller's inference block

is then defuzzified using the center of gravity method.

4.3 Question Exchange Requirement

As outlined in the content adaptation strategy in

Section 3.2, the Question Exchange Requirement

(QER) indicator plays a key role in determining

whether a student should be presented with questions

previously covered in the course during the final

assessment. This value is computed using Mamdani’s

expert fuzzy controller, which processes two input

variables: the student's competence in fundamental

machine learning concepts (competence – calculated

by the fuzzy controller described in Section 4.1 –

Machine Learning Competence controller) and the

student’s disengagement (disengagement – also

determined by a fuzzy controller, as discussed in

Section 4.2).

In line with the approach used for the fuzzy

controller calculating competence in machine

learning fundamentals, it was deemed essential to

keep the rule base for calculating the QER indicator

as minimal and interpretable as possible for the

expert. Consequently, the controller is based on two

input variables, each defined using three linguistic

values: low, medium, and high. These interpretations

are consistent across both variables, as shown in

Figure 6. Following the Mamdani model, the output

variable in the defuzzification block is also defined

using three linguistic values, as depicted in Figure 7.

The definitions of proposed terms align with the

course’s grading criteria just as with Machine

Learning Competence fuzzy controller.

The next phase in developing the fuzzy controller

involved defining the rule base, which is presented in

Table 4. This table outlines all the rules employed by

the controller. In constructing these rules, it was

assumed that the Question Exchange Requirement

(QER) is primarily influenced by the student’s

disengagement. It was recognized that a student’s low

competence in fundamental machine learning

concepts may not necessarily stem from low

engagement with the course, and therefore, such a

student should not be monitored by the system to the

same extent as a student who is actually disengaged

or has low activity in the course. For instance, in

Rules 1 and 2, the QER indicator is classified as low,

accompanied by a low level of disengagement,

despite variations in competence levels. A similar

pattern is observed in the pairs of Rules 5 and 6, as

well as Rules 8 and 9, indicating that disengagement

has a more significant impact on the QER indicator

than the level of machine learning competence.

Figure 6: Model of the input variables values in the

controller assessing QER indicator.

Figure 7: Model of the output variable values in the

controller assessing QER indicator.

As with the previously described fuzzy

controllers, the antecedents of the rules connect the

two input variables through a conjunction (and). Both

the conjunction and the implication operator (then)

are implemented using the minimum operation. The

fuzzy set resulting from the evaluation of all the rules

for specific input values is then defuzzified using the

center of gravity method.

CSEDU 2025 - 17th International Conference on Computer Supported Education

102

Table 4: Fuzzy controller rule base for calculating basic

machine learning competence.

No. Rule

if disengagement is low and competence is

h then QER is low

if disengagement is low and competence is

medium then QER is low

if disengagement is low and competence is

low then QER is mediu

if disengagement is medium and

competence is hi

h then QER is low

if disengagement is medium and

competence is medium then QER is

mediu

if disengagement is medium and

competence is low then QER is mediu

if disengagement is high and competence is

h then QER is mediu

if disengagement is high and competence is

medium then QER is hi

if disengagement is high and competence is

low then QER is hi

4.4 Modelling Student Behaviors Using

Fuzzy Logic

Fuzzy logic enables the design of inference systems

capable of handling discontinuities and non-

linearities in decision-making, thereby more closely

approximating human-like reasoning more closely

than traditional logic systems. This approach results

in a significantly streamlined rule base, as

demonstrated in the present solution, where each rule

base contains only nine rules that are easy to

understand for educators. Expecting teachers to create

precise rules that capture student behavior using

specific numerical thresholds would be highly

impractical, as such rigidly defined rules would

struggle to accommodate the diversity of student

work patterns. Fuzzy logic addresses this challenge

by providing a flexible, adaptive framework that

simplifies rule configuration and customization,

making it more intuitive and effective to model

diverse student behaviors.

5 AI TUTOR EVALUATION

The "Introduction to Machine Learning" course

described in Section 3 was a part of the Artificial

Intelligence class, with 89 students participating. The

course was designed to be completed within six days,

culminating in a final test. As outlined in the content

adaptation strategy in Section 3, once a student’s

calculated competence level in fundamental machine

learning concepts (Machine Learning Competence)

reaches a high threshold (at least 0.8), the system

grants the student unrestricted access to all course

content. Otherwise, sequential access to course

components is maintained.

Table 5 presents the representative sample of the

collected data about the performance of the fuzzy

controller for calculating Machine Learning

Competence and illustrates the decision made by the

system based on the content adaptation strategy

outlined in Section 3.2. The first column of the table

represents the initial knowledge level of each student,

quantified by the score attained on the initial test. The

second column reflects the student's cumulative skill

acquisition, represented by the sum of points earned

on tests following each course module. The third

column displays the student's competence level, as

Table 5: Summary of normalized information about the

level of Machine Learning Competence (competence) in

fundamental machine learning concepts calculated with the

fuzzy controller based on two input variables: (1) the

student’s score on an initial diagnostic test (knowledge),

and (2) their scores obtained from assessments following

each course module unit (skill).

Knowledge Skill Competence Decision

0.20 0.95 0.64 Maintaining

sequential access

0.80 0.95 0.91 Open access to

all course

content

0.60 0.65 0.64 Maintaining

sequential access

0.80 0.25 0.23 Maintaining

sequential access

0.00 0.90 0.64 Maintaining

uential access

0.20 0.25 0.21 Maintaining

sequential access

0.40 1.00 0.63 Maintaining

sequential access

0.60 0.80 0.89 Open access to

all course

content

1.00 0.85 0.90 Open access to

all course

content

0.80 0.45 0.24 Maintaining

sequential access

0.40 0.80 0.63 Maintaining

uential access

0.60 0.85 0.90 Open access to

all course

content

AI Tutor: Adaptive e-Learning System Using Expert Fuzzy Controllers

103

determined by the fuzzy controller. The last column

outlines the system's decision regarding the

availability of unrestricted access to course content,

based on the student's computed competence value.

According to the adopted adaptation strategy, full

access is granted if the student's competence level

reaches or exceeds a threshold of 0.8. The majority of

students exhibited insufficient foundational

knowledge at the outset of the course, thereby

preventing the granting of unrestricted access to all

course materials.

According to the content adaptation strategy,

when the student is ready to take the final test, the

Question Exchange Requirement (QER) indicator is

computed. If the QER value is below 0.5, it indicates

that the student has demonstrated sufficient

engagement throughout the course and possesses a

solid level of competence in fundamental machine

learning concepts. In this case, the final test consists

of questions that were previously included in the

course module summary tests (preliminary set of

questions). Conversely, if the QER value is 0.5 or

higher, it suggests that the student may not have fully

engaged with the material, and thus, the final test will

feature other questions designed to assess the

student's acquired skills throughout the course

(alternative set of questions).

Table 6 presents a normalized sample of data

about student coursework performance in relation to

their level of disengagement which is needed to

calculate the QER indicator. The first column

represents the system's calculated learning quality,

while the second column indicates the remaining time

to complete the final test, measured from the moment

the student began the course. The third column

records the disengagement level as determined by the

fuzzy controller. This disengagement value is

subsequently used as input for a third fuzzy

controller, which computes the Question Exchange

Requirement (QER) indicator. Table 7 shows the

results of this process, with the first column

representing the calculated level of competence in

fundamental machine learning concepts, derived

from the first fuzzy controller, and the second column

indicating the disengagement level. The third column

displays the QER value calculated by the fuzzy

controller, and the fourth column shows the system's

decision regarding the selection of questions for the

final test, based on the content adaptation strategy

outlined in Section 3.2. When a student's

disengagement level reaches at least medium-high, an

alternative set of questions is almost certainly

selected, unless the student's level of competence in

fundamental machine learning concepts is

sufficiently high and the student’s disengagement

level is low or medium.

Table 6: Summary of normalized information about the

level of disengagement calculated with the fuzzy controller

on the values of Learning Quality and Remaining Time.

Learning

Qualit

Remaining

Time

Disengagement

0.59 0.49 0.50

0.75 0.76 0.26

0.70 0.60 0.38

-0.16 0.01 0.84

-0.04 0.29 0.64

0.09 0.91 0.50

0.53 0.78 0.21

0.60 0.06 0.50

0.82 0.26 0.43

-0.09 0.55 0.50

0.82 0.66 0.17

0.11 0.06 0.84

Table 7: Summary of normalized information the Question

Exchange Requirement indicator calculated with the fuzzy

controller on the values of Machine Learning Competence

(Competence) and Disengagement.

Competence Disengagement QER Set of

questions in

the final test

0.64 0.50 0.52 Alternative

0.91 0.26 0.22 Preliminary

0.64 0.38 0.52 Alternative

0.23 0.84 0.90 Alternative

0.64 0.64 0.56 Alternative

0.21 0.50 0.64 Alternative

0.63 0.21 0.23 Preliminary

0.89 0.50 0.20 Preliminary

0.90 0.43 0.20 Preliminary

0.24 0.50 0.63 Alternative

0.63 0.17 0.21 Preliminary

0.90 0.84 0.64 Alternative

The fuzzy controllers implemented in the course

were designed to reflect the knowledge and

experience of the teachers responsible for the classes

where the course was introduced. Thus, the results

achieved met the expectations of the educators in

terms of providing different sets of questions

depending on the diagnosed level of student

engagement in learning. However, ensuring the

effectiveness of these controllers requires an in-depth

didactic-psychological study. Such research is

CSEDU 2025 - 17th International Conference on Computer Supported Education

104

essential because student behavior during the course

is influenced by various factors, including individual

student characteristics, such as their preferred

learning style, and different didactic conditions, such

as work overload, varying levels of interest in the

course topics, or even personal circumstances, such

as health challenges. Moreover, such a study is

challenging because students engage with the course

in an asynchronous mode without direct teacher

supervision. Despite these difficulties, this study is

planned for the future. It is expected that, given the

flexibility afforded by the asynchronous format, a

larger proportion of students will complete the course

and be evaluated using test questions from the

alternative set rather than the preliminary set,

allowing for a more comprehensive assessment of the

impact of the course.

6 CONCLUSIONS

The development of personalized e-learning systems

has become a focal point of modern education. The

AI Tutor system, introduced in this paper, is designed

to create adaptive content for the e-learning course

“Introduction to Machine Learning” providing

individualized learning experiences. This system

utilizes expert fuzzy controllers, a key technology

that is particularly effective in situations where

learning data is imprecise or uncertain. By leveraging

human-understandable if…then rules, the fuzzy

controllers help guide students through the course

material based on their interaction patterns, allowing

for dynamic content adaptation.

Expert fuzzy controllers differ significantly from

traditional machine learning methods. They do not

require datasets for training, making them an

attractive option in environments where data may be

sparse or difficult to model. Instead, these controllers

rely on expert knowledge, typically drawn from the

experience of teachers, to generate the rules that drive

the adaptation of learning content. These rules are

framed in terms of imprecise concepts reflecting the

nuanced nature of learning that is difficult to quantify

precisely.

The advantage of expert fuzzy controllers is their

ability to handle such imprecision effectively. As a

result, they can adapt the course content based on a

student's progress and behavior without the need for

training cycles. This characteristic makes them well-

suited for online education, where individual learning

paths can vary significantly. However, this approach

does have limitations. One major challenge is that the

rule base of the fuzzy controllers may not encompass

all possible patterns of student behavior. As a result,

there may be scenarios where the controller fails to

adapt the content appropriately or misses subtle

variations in how students engage with the course

material. This limitation highlights the inherent trade-

off between the simplicity and flexibility of expert

knowledge versus the complexity and adaptability of

machine learning-based approaches.

Another potential limitation of the AI Tutor

system is the risk that students might attempt to

circumvent the system. For instance, students could

collaborate on solving tests, leading to discrepancies

in how well the system reflects individual learning

progress. Although this issue does not undermine the

system's ability to personalize content, it points to a

need for ongoing analysis of student interactions. In

the future, a more in-depth examination of collected

data could provide insights into whether such

behaviors are widespread and how they impact the

overall effectiveness of the system. Addressing this

issue will be crucial in refining the system’s capacity

to adapt to diverse student behaviors and ensure that

it remains an accurate reflection of individual

learning experiences.

Looking ahead, there are opportunities to enhance

the AI Tutor system by exploring the automatic

generation of fuzzy controllers based on collected

data. By analyzing student interaction patterns over

time, it may be possible to improve the precision of

the controllers, either by better modeling the variables

used or by identifying additional patterns in student

behavior that should be included in the rule base.

Alternatively, machine learning models could be

trained on the same data and compared with the

expert-driven fuzzy controllers. This comparison

could shed light on the relative strengths and

weaknesses of both approaches, allowing for future

improvements to the system.

In conclusion, expert fuzzy controllers provide a

promising and effective solution for content

adaptation in e-learning environments. By leveraging

expert knowledge in a human-understandable form,

these controllers can dynamically tailor course

content to the individual needs of students. However,

the approach does have certain limitations,

particularly related to the coverage of all student

behavior patterns and the potential for students to

circumvent the system. Future work will focus on

refining the fuzzy controllers, exploring the

integration of machine learning techniques, and

addressing behavioral concerns to enhance the

system's effectiveness and adaptability.

AI Tutor: Adaptive e-Learning System Using Expert Fuzzy Controllers

105

REFERENCES

Böcker, H.-D., Hohl, H., and Schwab, T. (1990). Upsilon-

pi-ADAPT-epsilon-rho: Individualizing hypertext.

Proceedings of the IFIP TC13 Third Interational

Conference on Human-Computer Interaction, 931–

936.

Caro, M. F., Josyula, D., and Jiménez, J. (2015). Multi-level

pedagogical model for the personalization of

pedagogical strategies in intelligent tutoring systems.

Dyna, 82, 185–193.

Casillo, M., Conte, D., Lombardi, M., Santaniello, D., and

Valentino, C. (2021). Recommender System for Digital

Storytelling: A Novel Approach to Enhance Cultural

Heritage. In A. Del Bimbo, R. Cucchiara, S. Sclaroff,

G. M. Farinella, T. Mei, M. Bertini, H. J. Escalante, and

R. Vezzani (Eds.), Pattern Recognition. ICPR

International Workshops and Challenges (pp. 304–

317). Springer International Publishing.

Chandrasekhar, U., and Khare, N. (2021). An intelligent

tutoring system for new student model using fuzzy soft

set-based hybrid optimization algorithm. Soft

Computing, 25(24), 14979–14992. https://doi.org/

10.1007/s00500-021-06396-8

del Puerto Paule Ruiz, M., Díaz, M. J. F., Soler, F. O., and

Pérez, J. R. P. (2008). Adaptation in current e-learning

systems. Computer Standards & Interfaces, 30(1), 62–

70. https://doi.org/10.1016/j.csi.2007.07.006

Desai, D. (2022). Hyper-Personalization: An AI-Enabled

Personalization for Customer-Centric Marketing. In S.

Singh (Ed.), Adoption and Implementation of AI in

Customer Relationship Management (pp. 40–53). IGI

Global. https://doi.org/10.4018/978-1-7998-7959-

6.CH003

Dixson, M. D. (2015). Measuring Student Engagement in

the Online Course: The Online Student Engagement

Scale (OSE). Journal of Asynchronous Learning

Networks, 19. https://doi.org/10.24059/olj.v19i4.561

Dorça, F. A., Carvalho, V. C., Mendes, M. M., Araújo, R.

D., Ferreira, H. N., and Cattelan, R. G. (2017). An

Approach for Automatic and Dynamic Analysis of

Learning Objects Repositories through Ontologies and

Data Mining Techniques for Supporting Personalized

Recommendation of Content in Adaptive and

Intelligent Educational Systems. 2017 IEEE 17th

International Conference on Advanced Learning

Technologies (ICALT), 514–516. https://doi.org/

10.1109/ICALT.2017.121

Eduexe - eLearning authoring tool. (2024). Retrieved

November 16, 2024, from https://eduexe.com/

El-Sabagh, H. A. (2021). Adaptive e-learning environment

based on learning styles and its impact on development

students’ engagement. International Journal of

Educational Technology in Higher Education, 18(1),

53. https://doi.org/10.1186/s41239-021-00289-4

Ennouamani, S., and Mahani, Z. (2017). An overview of

adaptive e-learning systems. 2017 Eighth International

Conference on Intelligent Computing and Information

Systems (ICICIS), 342–347. https://doi.org/10.1109/

INTELCIS.2017.8260060

Fenza, G., Orciuoli, F., and Sampson, D. G. (2017).

Building Adaptive Tutoring Model Using Artificial

Neural Networks and Reinforcement Learning. 2017

IEEE 17th International Conference on Advanced

Learning Technologies (ICALT), 460–462.

https://doi.org/10.1109/ICALT.2017.124

Fleming, N. D. (2006). VARK visual, aural/auditory,

read/write, kinesthetic. New Zealand: Bonwell Green

Mountain Falls.

Gaudioso, E., Montero, M., and Hernandez-Del-Olmo, F.

(2012). Supporting teachers in adaptive educational

systems through predictive models: A proof of concept.

Expert Systems with Applications, 39(1), 621–625.

https://doi.org/10.1016/J.ESWA.2011.07.052

Google Teachable Machine. (2019). Retrieved November

16, 2024, from https://teachablemachine.withgoog

le.com/

Kasinathan, V., Mustapha, A., and Medi, I. (2017).

Adaptive learning system for higher learning. 2017 8th

International Conference on Information Technology

(ICIT), 960–970. https://doi.org/10.1109/ICITECH.20

17.8079975

Katsaris, I., and Vidakis, N. (2021). Adaptive e-learning

systems through learning styles: A review of the

literature. Advances in Mobile Learning Educational

Research, 1(2), 124–145. https://doi.org/10.25082/

AMLER.2021.02.007

Khomeiny, A. T., Restu Kusuma, T., Handayani, A. N.,

Prasetya Wibawa, A., and Supadmi Irianti, A. H.

(2020). Grading System Recommendations for

Students using Fuzzy Mamdani Logic. 2020 4th

International Conference on Vocational Education and

Training (ICOVET), 1–6. https://doi.org/10.1109/

ICOVET50258.2020.9230299

Kovacic, Z., and Bogdan, S. (2018). Fuzzy controller

design: theory and applications. CRC press.

Lerís, D., Sein-Echaluce, M. L., Hernández, M., and Bueno,

C. (2017). Validation of indicators for implementing an

adaptive platform for MOOCs. Computers in Human

Behavior, 72, 783–795. https://doi.org/10.1016/j.chb.

2016.07.054

Mamdani, E. H. (1974). Application of fuzzy algorithms for

control of simple dynamic plant. Proceedings of the

Institution of Electrical Engineers, 121(12), 1585–

1588.

Marciniak, J. (2014). Building e-learning content

repositories to support content reusability.

International Journal of Emerging Technologies in

Learning, 9(3), 45–52. https://doi.org/10.3991/ijet.v9

i3.3456

Marciniak, J., Szczepański, M., Dyczkowski, K.,

Mazurowska, K., Stanczewski, R., Grzybek, J., and

Marciniak, D. (2023). The Use of a Fuzzy Rule-Based

System in Adaptive e-Learning Content Based on

Intercultural Competence. 2023 IEEE International

Conference on Fuzzy Systems (FUZZ), 1–6.

https://doi.org/10.1109/FUZZ52849.2023.10309813

Mendel, J. M. (2017). Uncertain Rule-Based Fuzzy

Systems. Uncertain Rule-Based Fuzzy Systems.

https://doi.org/10.1007/978-3-319-51370-6

CSEDU 2025 - 17th International Conference on Computer Supported Education

106

Premlatha, K. R., and Geetha, T. V. (2015). Learning

content design and learner adaptation for adaptive e-

learning environment: a survey. Artificial Intelligence

Review, 44(4), 443–465. https://doi.org/10.1007/s104

62-015-9432-z

Raj, N. S., and V G, R. (2019). A Rule-Based Approach for

Adaptive Content Recommendation in a Personalized

Learning Environment: An Experimental Analysis.

2019 IEEE Tenth International Conference on

Technology for Education (T4E), 138–141.

https://doi.org/10.1109/T4E.2019.00033

Santos, P. B., Bhowmik, C. V., and Gurevych, I. (2020).

Avoiding bias in students’ intrinsic motivation

detection. International Conference on Intelligent

Tutoring Systems, 89–94.

Sushama, C., Arulprakash, P., Kumar, M., and Sujatha, K.

(2022). The Future of Education: Artificial Intelligence

based Remote Learning. International Journal of Early

Childhood Special Education, 14, 3827–3831.

Szczepański, M., and Marciniak, J. (2023). Application of

a fuzzy controller in adaptive e-learning content used to

evaluate student activity. Procedia Computer Science,

225, 2526–2535. https://doi.org/10.1016/j.procs.20

23.10.244

Vinaykarthik, B. C., and Mohana. (2022). Design of

Artificial Intelligence (AI) based User Experience

Websites for E-commerce Application and Future of

Digital Marketing. 3rd International Conference on

Smart Electronics and Communication, ICOSEC 2022

- Proceedings, 1023–1029. https://doi.org/10.1109/

ICOSEC54921.2022.9952005

Zadeh, L. A. (1965). Fuzzy sets. Information and Control,

8(3), 338–353. https://doi.org/10.1016/S0019-9958(65)

90241-X

AI Tutor: Adaptive e-Learning System Using Expert Fuzzy Controllers

107