IAES International Journal of Artificial Intelligence (IJ-AI)

Vol. 10, No. 1, March 2021, pp. 9~23

ISSN: 2252-8938, DOI: 10.11591/ijai.v10.i1.pp9-23 9

Design and analysis of a multi-agent e-learning system using

prometheus design tool

Kennedy E. Ehimwenma, Sujatha krishnamoorthy

Department of Computer Science, College of Science and Technology, Wenzhou-Kean University, China

Article Info ABSTRACT

Article history: Agent unified modeling languages (AUML) are agent-oriented approaches
that supports the specification, design, visualization and documentation of an
Received Apr 5, 2020 agent-based system. This paper presents the use of prometheus AUML
Revised Dec 31, 2020 approach for the modeling of a pre-assessment system of five interactive
Accepted Jan 6, 2021 agents. The pre-assessment system, as previously reported, is a multi-agent-
based e-learning system that is developed to support the assessment of prior
learning skills in students so as to classify their skills and make
Keywords: recommendation for their learning. This paper discusses the detailed design
approach of the system in a step-by-step manner; and domain knowledge
Agent methodology abstraction and organization in the system. In addition, the analysis of the
Computational education data collated and models of prediction for future pre-assessment results are
First-order logic also presented.
Pre-assessment classification
Requirements engineering This is an open access article under the CC BY-SA license.

Corresponding Author:
Kennedy E. Ehimwenma
Department of Computer Science
Wenzhou-Kean University
88 Daxue Rd, Ouhai District, Wenzhou, Zhejiang, China
Email: [email protected]

1. INTRODUCTION
An agent software methodology is a set of guidelines that covers the entire life-cycle of a multi-
agent development process. A multi-agent system (MAS) is a system of interactive agents or autonomous
program modules. In general, a unified modelling language (UML) assists software developers to specify,
design, visualize and document software engineering processes that meets application requirements [1]. A
UML allow models to be created, considered, developed, and processed in a standard way from the initial
phase of analysis to design and implementation [2]. Systems implementation is focused on users’ needs as
well as system functionality with requirements specification as the driver. From start to finish, effective and
efficient system evolves from user interaction and the incremental principle of development. Software
development stages have shared abstraction in both object-oriented programming (OOP) methodology and
agent-oriented software engineering (AOSE). In OOP paradigm, these stages are: requirements gathering,
analysis, design, implementation, testing and maintenance. Whilst the AOSE process subsumes the steps in
OOP methodologies, the concepts for developing objects (in OOP) are however different from those in agent-
based systems. For instance, object-oriented methodologies cover concepts such as objects, classes and
inheritance. But in AOSE, design concepts are terms that view agents as autonomous, situated, reactive, and
social. This paper is a presentation of the application of prometheus [3, 4] agent-oriented methodology for the
static and dynamic design of an elearning MAS. Though there are several AOSE methodologies for
designing agent-based systems, the choice of prometheus was predicated on its structured and detailed step-
by-step procedure that supports how requirement statements can be acquired. The purpose of the system is to
pre-assess students’ prior learning, classify their skills, and make recommendation for appropriate material

Journal homepage: http://ijai.iaescore.com

10 ISSN: 2252-8938

suitable to their needs. Thus, the contribution of this paper is: i) To demonstrate requirements analysis and
design specifications for the development of an e-learning pre-assessment system using MAS. ii) To analyse
the descriptive functions and roles of multi-agents within an e-learning pre-assessment system. iii) To show a
detailed model of software engineering with agent UML (AUML) tool for teaching and learning. iv) To
demonstrate inter-agent communication for the assessments and classification of students' prior skill-set. v)
To analyse the data collated from the system using regression models of prediction. This paper continues
with the background logic of knowledge engineering for the system in which an abstract model of an
ontology tree traversal is discussed as applicable in the MAS implementation. In section 2, the paper presents
AUML tools and agent software development life cycle (ASDLC). In section 3, models of analysis and
design from the use of the prometheus design tool (PDT) is presented. Section 4 looks into implementation,
issues at experimentation, data collection and analysis; and section 5 is conclusion.

Logic program in decision support system

Logic formulas are formal specifications that are readily used to represent facts, statements and
propositions. Such formalism e.g., Horn clauses, answer set programming, first order logic formulas and
description logics are used to for reasoning-supported decision processes in a dynamic system. With logic
programs and its diverse variants of formalization, facts and objects have been collected, categorized, and
relations established in-between objects of facts; and decisions taken e.g., [5, 6]. Approaches for MAS
development using first order logic (FOL) have also been demonstrated in [7, 8]. In a hybrid distributed
system in which asymptotic consensus result was obtained, [7] presented a leader-follower consensus MAS
in which the leader system shared knowledge with the follower system whose description was given in FOL.
Similarly, recent studies in MAS e.g. [8-11] have emphasized the need for adaptive elearning systems that
can personalize learning so as to meet individual learner needs. This is because what a learner wants, may
actually, be different from what he needs to learn. This research addresses this gap in the development of
elearning systems.

Background logic of knowledge engineering for the pre-assessment system

Let 𝑝 be a predicate. A binary relation between objects 𝑥1 and 𝑥2 can be given symbolically as
𝑝(𝑥1 , 𝑥2 ). Also, let ⅅ be a domain of directed ontological nodes, and i, j, and k = 1, 2, 3, …, n, n + 1,
respectively. If i represents the level of hierarchies of nodes on the horizontal traversal, and j the
arrangements of nodes on the vertical traversal and 𝑝𝑖,𝑗 some random predicates or property such that 𝑥𝑖,𝑗 is a
parent node, 𝑥𝑖+1,𝑗+1 a perquisite parent node next to 𝑥𝑖,𝑗 , and 𝑧𝑗,𝑘 a leafnode; then the following abstraction
holds:
Ɐx ϵ ⅅ, 𝑝𝑖,𝑗 (𝑥𝑖,𝑗 ), which states that, every node has a property.
Ɐx ϵ ⅅ, 𝑝𝑖,𝑗 (𝑥𝑖,𝑗 , 𝑥𝑖+1,𝑗±1 ), which means a parent to prerequisite parent node relation. Any parent node
is a desired topic to be learned by students.
Ɐx ϵ ⅅ, 𝑝𝑖+1,𝑗 (𝑥𝑖,𝑗 , 𝑧𝑗,𝑘 ), that there is a direct relation of a parent node to its own leafnodes.
ⱯxⱯz ϵ ⅅ, 𝑝𝑖,𝑗 (𝑥𝑖,𝑗 , 𝑥𝑖+1,𝑗±1 ) ꓥ 𝑝𝑖+1,𝑗±1 (𝑥𝑖+1,𝑗±1 , 𝑧𝑗±1,𝑘 ) → 𝑝𝑖+1,𝑗±1 (𝑥𝑖,𝑗 , 𝑧𝑗±1,𝑘 ), which is a transitive
relation for navigating leafnodes connected to prerequisite parent nodes, or
ⱯxⱯz ϵ ⅅ, 𝑝𝑖,𝑗 (𝑥𝑖,𝑗 , 𝑥𝑖+1,𝑗±1 ) ꓥ 𝑝𝑖+1,𝑗±1 (𝑥𝑖+1,𝑗±1 , 𝑧𝑗±1,𝑘 ) → 𝑝𝑖,𝑗±1 (𝑥𝑖,𝑗 , 𝑧𝑗±1,𝑘 ), which are transitive
relations for traversals of leafnodes connected to parent nodes.
With a tree diagram as shown in Figure 1, the above stated symbolic relations are further deduced as follows:
That the parent nodes or the objects of learning are the c1, c2, c3, c4, c5, and c6. Along the horizontal
traversal, Level 1 to Level 4, the preceding node e.g., c1 is a parent to c2 and c3; and c2 and c3 are in turn
perquisites nodes to c1. Then, c2 and c3 are parent nodes to c4 and c5, respectively; and c4 and c5 are in turn
prerequisites to c2 and c3, respectively. Further down, c4 is a parent to c6, and c6 in turn is a prerequisite to
c4. In the tree, the set of 𝑐 ∈ C have their respective leafnodes 𝑁𝑖 labelled as N1, N2, N3, …, N11 are leafnodes
𝑛 ∈ 𝑁.

Establishing navigational relations between nodes

On the basis of the tree as shown in Figure 1, we now show the relationship between nodes in the
tree and the symbolic axioms stated earlier as implemented on the multi-agent pre-assessment system that:
𝑝𝑖,𝑗 (𝑥𝑖,𝑗 ) describes a unary predicate, and an example is desiredConcept(c1); and 𝑝𝑖,𝑗 (𝑥𝑖,𝑗 , 𝑥𝑖+1,𝑗±1 ) describes
a binary relation which states that a parent has a named prerequisite, and an example of the form is
hasPrerequisite(c1, c2); and 𝑝𝑖+1,𝑗 (𝑥𝑖+1,𝑗 , 𝑧𝑗,𝑘 ) which states that a parent has a named leafnode, and an
example is formula hasKB(c6, n11). Then for all pre-assessment and recommendation of any failed learning
unit i.e., the leafnodes N, we state that

Int J Artif Intell, Vol. 10, No. 1, March 2021: 9 – 23

Int J Artif Intell ISSN: 2252-8938 11

𝑝𝑖,𝑗 (𝑥𝑖,𝑗 , 𝑥𝑖+1,𝑗±1 ) ꓥ 𝑝𝑖+1,𝑗±1 (𝑥𝑖+1,𝑗±1 , 𝑧𝑗±1,𝑘 ) → 𝑝𝑖+1,𝑗±1 (𝑥𝑖,𝑗 , 𝑧𝑗±1,𝑘 ). axiom 1

Axiom 1 states that if a parent node 𝑥𝑖,𝑗 has a named prerequisite 𝑥𝑖+1,𝑗±1 (one level below the hierarchy) of
𝑥𝑖,𝑗 either on its right hand or left-hand side which is denoted by 𝑗 ± 1 (+ for right, and – for left), and the
named prerequisite has a named leafnode 𝑧𝑗±1,𝑘 , then the parent node has a direct relation with the leafnode
like the perquisite has. An example of this transitive closure is hasPrerequisite(c4, c6) ꓥ hasKB(c6, n11) →
hasKB(c4, n11) that satisfies the property of transitivity. In addition, the axiom 1 conveys the leafnodes 𝑧𝑗±1,𝑘
that are: i) pre-assessed upon, and ii) the nodes that are recommended when any leafnode N that is connected
to the prerequisite node 𝑥𝑖+1,𝑗±1 are failed. On the other hand, the counterpart axiom 2

𝑝𝑖,𝑗 (𝑥𝑖,𝑗 , 𝑥𝑖+1,𝑗±1 ) ꓥ 𝑝𝑖+1,𝑗±1 (𝑥𝑖+1,𝑗±1 , 𝑧𝑗±1,𝑘 ) → 𝑝𝑖+1,𝑗 (𝑥𝑖,𝑗 , 𝑧𝑗±1,𝑘 ) axiom 2

is the axiom that also satisfies the property of transitivity. In this case, it is for the recommendation of
leafnodes 𝑧𝑗±1,𝑘 that has direct relations to the desired topic’s 𝑥𝑖,𝑗 given that an episode of pre-assessment on
the perquisite 𝑧𝑗±1,𝑘 connected to 𝑥𝑖+1,𝑗±1 have all been attempted and are all passed. An example of this
logical axiom 2 is hasPrerequisite(c1, c3) ꓥ hasKB(c3, {n4,n5, n6}) → hasKB(c1, n1).

Figure 1. A knowledge graph of multiple horizontal and vertical traversal

In our agent-based pre-assessment system, agents need to communicate the ground fact representation of this
logical axioms. For instance, for an agent to resolve the relevant plans for their next action, this group of
agents must inter-communicate the desired topics, passed leafnodes and/or failed leafnodes using the
following predicate logic form passed(𝑧𝑗+1,𝑘 ) and failed(𝑧𝑗±1,𝑘 ). The predicates which are the actions taken
by the multi-agents based on students’ response to questions, form the basis of the facts about the outcome of
a student using logic programming. This is because any object has a property that it satisfies or that any
object is connected by some relation to another object. From the foregoing, the explicitly stated logic-based
formulas are the premises on which the multi-agent of the pre-assessment system interacts, select ontological
nodes, select questions associated with leafnodes, assess users, classify user skills and recommend learning
materials.

2. RELATED WORK ON AUML AND E-LEARNING SYSTEMS

A UML and AUML diagrams are architectural model for information systems development. AUML
in particular, has been used in the design and specification of systems in the areas of weather forecasting,
business trading, petri-net, miner robot; and agent negotiation before implementation. In the field of
educational systems, agent-based system development research on intelligent agent models for elearning
have received little attention. Amongst this few are the works of [12] and [13]. In [12], Gaia methodology
was applied in the analysis and design of an elearning based MAS. The system which was proposed for
implementation on JADE framework is a security-based MAS that was meant to detect threats and provide
protection on web-based learning management systems (LMS) such as Moodle. In [13] an analysis and
design of a web-based services for virtual learning environment (VLE) was also discussed. The paper
Design and analysis of a multi-agent e-learning system using prometheus… (Kennedy E. Ehimwenma)
12 ISSN: 2252-8938

conceptualized a VLE application on mobile agent technology for the assessment of students’ knowledge,
and described agent role and agent interaction using a UML tool, and finally to implementation using JADE.
The InfoStation system [2] is a project of distributed elearning centre (DeLC), also used multi-agent
technology with proposed implementation on JADE [14]. With a UML, [2] described the InfoStation system
as a system of interactive agents whose functions included designated e-services. Also, in [15] the AGILE-
PASSI methodology was reported as the development tool for a medical educational game called
MEDEDUC for the purpose of improving learning in medical education and clinical performance. As a,
MEDEDUC allowed students to answer questions at different level of difficulty on multimedia presentation.
While many applications on agent-based technology are developed in fields such as commerce and security,
or adaptive dynamic programming [16] very limited attention has been given to agent-based development for
student learning. Among the aforementioned few, none had the combined system goal of skills classification
and recommendation of learning materials that we are presenting in this paper.

2.1. Agent-oriented methodology

Methodologies as a process of engineering a software enables developers to concretise the various
interaction components of a system and the functions needed amongst the various components for a system to
be realised. The work of [17] acknowledged that several AOSE methodologies exists for the analysis and
design of MAS but there also exists difficulty in the choice of the appropriate methodology for software
solutions from domain to domain. For instance, amongst gaia [18-21], tropos [22, 23], MaSE [24], PASSI
[25, 26] and prometheus [27] methodologies what factor should inform the developer’s choice? Though these
methodologies show similarities in their design process, there are however a varying degree of differences:
From requirements analysis, through to functionality modelling for agents, and implementation. In the
following section, Promethous is presented; and in Table 1 is a comparative summary of Gaia, Tropos and
Prometheus methodologies with regards to their similarities and differences, and the basis upon which the
Prometheus methodology was choosing for the design of the pre-assessment system. As shown in Table 1,
one common similarity is the customised tool associated with each methodology to support their engineering
process. The difference in their respective engineering process can be found in their design steps. For
instance, the tropos concept of softgoals which is equivalent to subgoals in prometheus is a breakdown of
hardgoals and initial goal of agents (or actors) functionalities, respectively. This, [28] referred to as role
decomposition which reduces the complexity in MAS engineering. The basis for prometheus is the design
step for deriving initial goals.

Table 1. A comparative summary of gaia, tropos and prometheus

Methodologies Phases Comparison
* Lack detailed step-by-step breakdown.
* Statement of requirement * No details on how requirement statements may be acquired.
Gaia * Analysis * View agent system as an organisational model.
* Design * Roles are similar to functionalities in Prometheus.
* Editor tool Gaia4E supports design.
* Emphasises the Early Requirement Analysis, then the Later Requirement Phase.
*Early requirement phase * Specialisation of Goals into subclasses of Hardgoal, and Softgoals for actors of
* Later requirement phase system.
Tropos * Architectural design * No general architecture containing all the phases of design as in Gaia, MaSE, or
*Detailed design Prometheus.
* Implementation * Has a design support tool called Taom4E.
* No Early Requirement phase as in Tropos. But this can be adapted.
* System specification * Uses Initial goals, that are refined or broken down into Subgoals for agents.
Prometheus * Architectural design * Very detailed design activity from System Specification phase to other phases.
* Detailed design phase * Reliance on expert knowledge on domain subject for requirement acquisition.
* Has a customised PDT, a AUML tool that supports design process.

2.2. Prometheus
Prometheus [27] is a methodology designed for the realisation of BDI agent systems with the use of
goals and plans. It supports development activities from requirements specification through to detailed design
for implementation. Prometheus design tool (PDT) [29, 30] is a graphical editor that supports the Prometheus
methodology. The PDT supports the development and documentation of all the phases of the Prometheus
methodology for building agent-based systems. Prometheus has three inter-connected design phases which
are system specification, architectural design, and the detailed design.

Int J Artif Intell, Vol. 10, No. 1, March 2021: 9 – 23

Int J Artif Intell ISSN: 2252-8938 13

3. MULTI-AGENT BASED PRE-ASSESSMENT SYSTEM DESIGN METHODOLOGY

In this section, we present the prometheus design methodology and detailed analysis of the pre-
assessment system. In Table 2 are the PDT symbols and description of their functions in the design of agent-
based systems. The pre-assessment system is a formative assessment system designed to supports the
learning of SQL. Learning as asserted in [31] is an effort intensive task. Thus, designing a MAS for learning
purposes is a complex process [28].

3.1. System specification

The system specification phase begins with a high-level description of the problem, which leads to
the identification of initial goals decomposition as shown in Figure 2.

Initial goals:
§ Observe percept Goal Specification:
§ Understanding of prerequisite § HOW CAN EACH OF THESE * Observe percept
§ Testing INITIAL GOALS BE ACHIEVED ? - Receive user concept
§ Classifying § E ACH GOALS HAD FURTHER - present concept
§ Continuous feedback SUB-GOALS DEVELOPED AS : DESIRED_CONCEPT
§ KB update
§ Recommend materials

*Testing * Understanding of prerequisite

- search BB for quizzes - quizzes in BB
*Classifying - fetch (sub-concepts or) prerequisite quizzes - answers in BB
- aggregate learning activity
- receive answer - prerequisite assessment from quizzes & answers
- use predicate statement rules UNDERSTANDING PREREQUISITE
- fetch BB answer and compare with students’
- classify students based on rules match
- make assessment decision
CLASSIFICATION TESTING USER

*Recommend materials
*KB updating - concept ontology in BB
*Continuous user feedback - search ontological relation
- store user learning activity persistently -user friendly interaction from assessments
PERSISTENT BELIEF STORE - fetch URL link
-welcome and introduction to system - present to user
USER INTERACTION
RECOMMENDATION

Figure 2. High level description of problem including initial goal and overall system goal specification

Table 2. The PDT notation symbol and meaning [3, 32, 33]
Name Symbol Description
Agent The agent symbols.
Action This is what the agent does that has effect on the environment or other agents.
Role This symbolizes roles or group of roles for agents.
Protocols specifies interaction between agents. Protocols are specified using textual
Protocol
notations that maps to AUML2.
This is used to represent the belief (internal knowledge model) or external data. It is where
Data
functionalities that transcends to agent read or write data or information.
Messages This is used to symbolize a message communication between agents.
BDI Messages This symbol is used to represent messages that updates the beliefs of agents.
Percept Represents the input coming from the environment to the agent.
This is an abstract description of a sequence of steps taken in the development of a system.
Scenario It is usually the initial step that starts for the breakdown of the “statement of problem” or
description of the problem to solve.
Goal It is the realizable target or achievement set for an agent.
Connection Arrows They are edges that connects entities (i.e. symbols) together.

3.1.1. Scenario overview

Scenarios and system goals are complementary. In the process of extracting the main goals from the
problem description, scenarios were developed as shown in Figure 3. In Figure 3 are the set of scenarios
derived from the specified goals using the PDT scenario overview diagram.

3.1.2. System goal diagram

The PDT system goal overview diagram enables the break-down or refinement of the set of derived
scenarios into units of achievable design steps. The Figure 4 is the system goal and subgoals design and the
interactions between them. The AND is a conjunction function which indicates that, at that level of design,
the agent must communicate with both classify and the persistentBB update goals after its decision-making
function. In Figure 4, at the user interface, percept goal is seen interacting with the understanding of
prerequisite goal which connects to the testing goal. Then to the make decision goal that is linking both the
Design and analysis of a multi-agent e-learning system using prometheus… (Kennedy E. Ehimwenma)
14 ISSN: 2252-8938

classify and persistentBB update goals after its decision-making function; and the classify goal further
connects the recommend material goal.

Figure 3. System scenario view

Figure 4. System goals specification and decomposition for

the pre-assessment system

3.2. Architectural design

At this phase, the role of the system for the purpose of pre-assessment has been conceptualized. The
needed number of agents and their descriptive names have been determined and included in the design. This
phase covers the system overall (static) structure using system overview diagram, and the description of the
dynamic behaviour of the system using interaction diagram and interaction protocols. In Figure 5 are the
identified roles that are needed within the role decomposition results into a role hierarchy from super-roles to
atomic roles (top-down direction) [28].
In this step, all the agents, their percepts, incoming messages, actions duly taken and interaction in
the design are presented in Figure 6. In the system overview diagram, the data (agent knowledge) that is
expected to used is coupled with the agents. In this design, the data are quizzes, answers to quizzes, and URL
data links for each of the leafnodes (sub-topics) in the ontology. The data is modelled as internal knowledge
or beliefs in the agents. Figure 6 also presents the five working agents of the system whose detail design are
illustrated in agent overview stage.

Figure 5. System role overview showing structured

functionalities
Figure 6. System overview diagram

3.3. Detailed design

This phase is focused on the description of responsibilities and capabilities of the internal structure
of the individual agent, and how each agent would achieve their task within the system. Diagrammatically,
these capabilities were realised at the agent overview stage shown below.

3.3.1. Agent overview

In this section, the list of agents in the system and individual agent internal details are presented.
This includes the agents’ plan models, notation for percepts and triggering_event, communication links and
inter-agent message description. At the agent overview stage, inherited interfaces from e.g., the system

Int J Artif Intell, Vol. 10, No. 1, March 2021: 9 – 23

Int J Artif Intell ISSN: 2252-8938 15

overview phase are adopted for specifying agents’ details. The inherited interfaces are the notation symbols
that appear greyish in colour.

Agent agInterface: In Figure 7 is a much refined and detailed design where CArtAgO artifact is the medium
that was used to get user input. The interface agent first creates the artifact in order to observe it. The
observed inputs are communicated as messages in agent plan (shown with the plan diagram or symbol) to
other agents e.g., the agent agSupport that is responsible for pre-assessing students.
Agent agSupport: This is the pre-test agent that is saddled with the task of questioning a user’s skills before
making recommendation, as shown in Figures 8-9. The agent agSupport uses its achievement goals for
navigation, from leafnode 𝑧𝑗,𝑘 to leafnode 𝑧𝑗,𝑘+1 in the hierarchy of concepts to retrieve quizzes which are
represented in predicate logic in its BB to test students’ skills. Using the answer percept received, it compares
and matches the given answer input with the predefined answer in its BB. Taking the decision for either a
passed or a failed predicate on every answer received, this agent also communicate all assessment activities,
namely: the decision reached per question, the questions asked, and communication of the answers received
to other agents in the MAS that needs to know. This agent also date and timestamp every learning activity.

Figure 8. Agent agSupport receiving the

Figure 7. Detailed overview of agent agInterface
desired_Concept percept and retrieving
quizzes using plans

Agent agModelling: This agent gets message percepts from agent agSupport for every leafnode (question
attached to a unit of learning) in the ontology whose pre-assessment has been completed. This agent uses the
percept (or information) it receives to match the pre-conditions in its plan context, and thereafter classify the
student’s skills. The category of information (in one plan) that is determined by this agent is communicated
to the next receiving agent (agMaterial) that will in turn send learning material to the student, as shown in
Figure 10.

Figure 9. Agent agSupport overview

Figure 10. The agent agModelling (classifier agent)

Agent agMaterial: Figure 11 is agent agMaterial that keeps the URLs links of learning material as an
ontology. The perfomative used in the message to this agent is “achieve”. On receiving the “achieve”
performative message from the classifier agent (after classification), the agent agMaterial then releases
learning materials for students to learn. These materials are dependent on the number of failed and passed
prerequisite assessments.
Agent agModel: This agent uses the Java TextPersistentBB class to store all the learning activities in the
system. The TextPersistentBB class was configured in the MAS at the point of declaration or creation of the
multi-agents project with the Mas2j [34] extension at the level of implementation. The activities stored are
Design and analysis of a multi-agent e-learning system using prometheus… (Kennedy E. Ehimwenma)
16 ISSN: 2252-8938

messages sent to the agent; and they include students’ desired topics, and answers to question (both correct or
incorrect) percept. As shown in Figure 12, the persistent beliefs are permanently stored in the system.

Figure 11. Agent agMaterial: The learning material

agent overview Figure 12. Agent agModel (student) overview

3.4. Logical analysis of pre-assessment

Previous reports on the pre-assessment system presented a pre-assessment mechanism [35, 32] and
the formalized logic model [36] for pre-skills testing, decision making, selection and recommendation of
learning materials. In general, on any given knowledge graph or ontology, the following symbolic algorithm
presents the underlying reasoning [32]:

ⅅ 𝐶𝑖 𝑁𝑖,𝑗 hasPrerequisite(ⅅ𝑖,𝑗 , 𝐶𝑖+1 ) ꓥ hasKB(𝐶𝑖+1,𝑗±1 , 𝑁𝑗±1,𝑘 )

[
: Ǝⅅ ꓥ passed(𝑁𝑗±1,𝑘 ) => learn(𝑁𝑗,𝑘 ) ꓥ hasKB(ⅅ𝑖,𝑗 , 𝑁𝑗,𝑘 )
else
: Ǝⅅ ꓥ Ǝfailed(𝑁𝑗±1,𝑘 ) => learn(𝑁𝑗±1,𝑘 ) ꓥ hasKB(𝐶𝑖+1,𝑗±1 , 𝑁𝑗±1,𝑘 )
]

of the multi-agent pre-assessment system whose software engineering design steps we have presented in the
preceding sections; such that, ⅅ is the desired concept (also called the desired topic) that subsumes some
prerequisites 𝐶𝑖 which further subsumes some leafnodes 𝑁𝑖,𝑗 . In description logic notation, it states 𝑁𝑖,𝑗 𝐶𝑖
ⅅ. In the system, the content of learning is in the domain of SQL (structured queried language) from which
topics – that we have called the DesiredConcept ⅅ are chosen and studied by students.
Now, let ⅅ = { 𝑐 ∈ ⅅ | 𝑝(𝑐) } and N = { 𝑛 ∈ 𝐶 | 𝑞(𝑛) }. ⅅ precedes 𝐶 in the hierarchy of concepts (or topics)
of learning such that the number of elements in D = C + 1. Then the set of topics otherwise known as
elements considered in the domain ⅅ is given as
ⅅ = {union, join, update, delete, insert, select};
and the set of all prerequisites 𝐶 underneath ⅅ is given as
𝐶 = {join, update, delete, insert, select};
and the set of all terminal leafnodes 𝑁 in 𝐶 and ⅅ, respectively, given as
N = {union, unionAll, selfJoin, fullOuterJoin, innerJoin, updateSelect, updateWhere, deleteSelect,
deleteWhere, insertSelect, insertValue, selectOrderBy, selectDistinct, selectWhere, selectAll }.

In education, teaching-learning is chronological and this forms the basis for the connection of a previous
learning to a new or ongoing learning. Let the relation 𝑅 be the set of connection between nodes i.e. a new
topic and a previously learned topic. Then we state that D and C, and C and N belong to some relations,
respectively; as shown in the following set A and B with regards to the given ontological node relationships
A = D x C ∈ R and, B = C x N ∈ R.
Symbolically, it holds that
∀𝑑 ∈ ⅅ ∀𝑐 ∈ 𝐶 ∀𝑛 ∈ 𝑁, 𝑅(ⅅ, 𝐶) ∧ 𝑅(𝐶, 𝑁)
In furtherance, the elements of the sets D, C and N thus satisfies the following definitions and their respective
properties 𝑝 and 𝑞 in predicate formulas
(d, c) = {𝑝 ∈ 𝑅 | 𝑝(𝑑, 𝑐)}, where the relation 𝑝 = hasPrerequisite; and
(c, n) = {𝑞 ∈ 𝑅 | 𝑞(𝑐, 𝑛)}, where the relation 𝑞 = hasKB.
Symbolically, the conjunction of the above given relation is
∀𝑑 ∈ ⅅ ∀𝑐 ∈ 𝐶 ∀𝑛 ∈ 𝑁, hasPrerequisite(d, c) ∧ hasKB(c, n).

Int J Artif Intell, Vol. 10, No. 1, March 2021: 9 – 23

Int J Artif Intell ISSN: 2252-8938 17

By the property of transitivity,

hasPrerequisite(d, c) ∧ hasKB(c, n) → hasKB(d, n).
For example, choose the update ∈ ⅅ. Its immediate prerequisite delete ∈ 𝐶. The node update has its
leadnodes n = {updateSelect, updateWhere}, and the node delete also has its leafnodes n = {deleteSelect,
deleteWhere}. From the foregoing,
hasPrerequisite(update, delete) ∧ hasKB(delete, {deleteSelect, deleteWhere}) → hasKB(update,
{deleteSelect, deleteWhere}).
Then, the conclusion becomes the units to be pre-tested on, and if any is failed; students are recommended
materials for the failed units to learn. As stated earlier, if no pre-test is failed, i.e. all pre-assessments are
passed; then the conclusion as in the following thus hold
hasPrerequisite(update, delete) ∧ hasKB(delete, {deleteSelect, deleteWhere}) → hasKB(update,
{updateSelect, updateWhere})
which are the leafnodes of the chosen topic. The stated logic-based representation, formed the ground fact
internal knowledge model of the multi-agent pre-ssessment system designed in this study.

4. DISCUSSION
The paper has presented the prometheus AUML design tool for the design and analysis of the pre-
assessment system, and its implementation with Jason – a Java-based interpreter and declarative language.
The choice of Prometheus methodology ensured that every requirement and detailed design activity were
captured with the appropriate symbol. This we have depicted from initial goal specifications, to subgoals, to
agent roles and interaction using distinctive diagrams. From critical analysis, Prometheus provides support
on how requirement statements may be acquired -- starting with intial goals specification -- as well as a
general system architecture as against some other AUML tools. These steps are vital as any left-out
functionality would cause a void in the system: A void that may require the re-engineering of the whole
system. In a declarative language, agents communicate via message passing in predicate logic form. Thus, in
line with the reported mechanism of pre-assessment & recommendation and formalized (FOL based) pre-
assessment rules [35-37] in which the MAS made accurate recommendation after pre-assessment, Figure 13
hereby presents the pseudocode of the operation of the system and how the perceived knowledge by agents
are used: from percept acquisition at the interface (line 7), through to other agents via the .send() internal
action [4] (on lines 9, 11, 18, 23, 27; as shown in Figure 13) which clearly shows the number of interactive
agents in the system. Between each internal
Pseudocode action,andisinteraction
of pre-assessment the action designated
in the for a receiving agent to execute.
multiagent system

1. initial beliefs: predicate(Class, Class)

2. initial beliefs: predicate(Class, Leafnode)
3. initial beliefs: predicate(Leafnode, URL)
4. initial beliefs: quiz(PrerequisiteLeafnode)
5. Given a desired concept that has N leafnodes prerequisite
6. IF
7. Percept ← desiredConcept
8. THEN
9. .send(receiver, tell, desiredConcept)
10. fetch the next quiz(Prerequisite_Leafnode)
11. .send(receiver, tell, quiz(Prerequisite_Leafnode)
12. output quiz(Prerequisite_Leafnode)
13. Percept ← answer(X)
14. IF
15. answer(X) == answer(Prerequisite_Leafnode)
16. THEN
17. passed(Prerequisite_Leafnode) decision
18. .send(receiver, tell, passed(Prerequisite_Leafnode)
19. IF
20. answer(X) \== answer(Prerequisite_Leafnode)
21. THEN
22. failed(Prerequisite_Leafnode) decision
23. .send(receiver, tell, failed(Prerequisite_Leafnode)
24. IF
25. N number of leafnodes have been pre-assessed on
26. THEN
27. .send(receiver, achieve, recommendMaterial)
28. Else
29. repeat 10 to 27

Figure 13. Pseudo-algorithm of the pre-assessment process that depends on the number of leafnodes N
considered under a desiredConcept

Design and analysis of a multi-agent e-learning system using prometheus… (Kennedy E. Ehimwenma)
18 ISSN: 2252-8938

4.1. Resolving issues of development

As a pre-assessment system for SQL pre-skills test and skills classification into two [1, 0] binary
states, the system receives open-ended SQL query inputs that maybe correct or incorrect answer for a
particular query question. For a given query, the inflow of incorrect answers to the system is not definite nor
predetermined compared to correct query answers that are known and predefined in the system. Thus,
programming a MAS for the recognition of negative facts (i.e. incorrect answers) can pose some difficulty
for agent plan selection and execution of agent goals when the expected inputs may vary depending on a
student understanding and query competences. In such cases, inputs become diverse, unbounded and
subjective to the students. On the hand, the correct SQL queries which are the positive facts are quite
straightforward to program because every answer to the questions asked on the system is predetermined
based on standard SQL queries. Typically, the syntax of Jason agent plan comprises three parts, and the
structure given as: triggering_event, context <-- body [4]. The context is the part of the plan which states the
pre-condition that activates a plan for execution. By default, a blank plan context is true for all beliefs in the
agent. Otherwise, a predicate form pre-condition must be stated to control what plan is right for a given
triggering_event and beliefs. To determine whether a query input is right or wrong, the predefined positive
facts were represented in the pre-assessment agent in first order logic (FOL) predicate form – for a
declarative language.
When an agent gets percept (also in FOL form), the agent matches that percept (now a belief)
against all plans, the relevant plan is selected and actions in the body of the plan is executed --- in One vs. All
approach. This is because positive facts are information whose representation are known and can be
represented or given to the agent (i.e. responsible for handling pre-skills assessment) for comparison with
incoming percepts. So, the correct SQL queries were initialized in the agent’s belief base and were used by
the agent to match and trigger relevant plans and agent goals as needed. But negative facts are unknown and
as such cannot be pre-determined for representation as mentioned earlier. So, to address incorrect SQL query
inputs, Jason different \== operator [4] was used as the comparison operator in the agent plan context. In
analogy, the operator means “not equal” or “false”. But the use of this operator was not without
inconsistency in the collective multi-agents behaviour during system coding and implementation phase.
During coding, when an incorrect query was inputted for testing, the \== operator made the pre-assessment
agent to miss-select plans from its plan library. For example, by stating in a plan that was expected to handle
an incorrect SQL query, that,

if the answer percept coming into the system does not match some initially predefined SQL
queries then inform the student that the answer given is incorrect and then select the next
leafnode question and present to the student.

Literally, from the behaviour exhibited by the agent, the agent’s interpretation was any other plan whose plan
context has no match to any already known knowledge in the agent’s belief base. The miss-selection of plans
was due to some uncertainty in the agent ability to map an incorrect query percept to beliefs. This behaviour
as observed adversely altered the order of subsequent goal/question selection of a prerequisite’s leafnode N,
in contrast to the arrangements of nodes in the ontology tree. This was a non-trivial problem. At the
implementation phase, one of the key principles of software methodology is to combine coding and testing
[38]. This principle which enables a system to be investigated while it is still being developed ensured that
this non-trivial problem was checked before the system was completely built.
To enable the pre-assessment agent, as shown in Figures 8-9 to accurately select relevant plan(s) for a match
of its plan context to the percept that is adopted in the \== operator; and to correctly determine the next
appropriate agent goal and accurate message passing to other agents, we had to introduce a process of
iteration that could count plan selection in the agent for every parent node (or topic) and their connected
leafnodes N. In Jason predicate logic form, an example syntax of this iteration is countForDeletePre(X),
which depicts the counter for the delete node where X is a positive integer. In addition, the negation of some
incoming percept was required to stop unsolicited plan trigger. An example of such negation in the context
part of a plan, was the not desiredConcept(“insert”) which was used to block-off the
desiredConcept(“insert”) in a plan context so as not trigger the wrong plan and wrong agent goal at a given
time. As the system kept expanding with the programming of more parent nodes D and leafnodes N been
added, this block-off continued and it used to mitigate agent anomaly behaviour. These two combined
strategies effectively controlled agent behaviour as well as the entire multi-agents toward handling of the
incorrect SQL query inputs.

Int J Artif Intell, Vol. 10, No. 1, March 2021: 9 – 23

Int J Artif Intell ISSN: 2252-8938 19

4.2. Experimentation and results

SQL programming has been reported to pose challenges to student learners e.g., [39]. Our previous
survey conducted in [32] also found that students are faced with SQL query difficulties. In this paper, we
have slightly varied the approach to prior-skills assessment that was taken in previous survey. This was to
determine whether changing any factor before and during the course of pre-assessment can cause any
significant change in the results obtained in previous researches. Against previous research, the two
approaches introduced are, namely: i) To allow students have adequate familiarity with the database (which
has five tables) prior to pre-skill assessments; and 2) To allow students consult with any text or written
materials throughout the time of their pre-skill assessments on the system. But with no prior access to the
questions attempted at their pre-skills evaluation. From the results obtained in this paper, changes in the
approach of pre-testing has slightly improved results, as shown in Tables 3-4 on student pre-skills exercise in
contrast previous surveys in [32, 39, 40]. As students used the system, the system stored all pre-assessment
activities and also timestamped every entry made by students into the system. Figure 14 is a screenshot of the
pre-assessment system interface that is presenting the Union desired topic, and subsequent interaction
between the student and the MAS in the course of pre-assessment. Agents have independent autonomy and
mental states. The MAS leveraged on these properties coupled with their interaction protocol [7] and we
obtained optimal recommendation of learning materials for either all passed or any failed pre-assessments.

Figure 14. Agent interface showing the Union desired concept

4.2.1. Agent mind inspection

Upon the agent agModel Mind Inspection, some examples of the structured beliefs stored by this
PersistentBB agent in FOL predicate form are:

which is the belief about a student’s desired topic, and

the persistence belief of the student’s response to a pre-assessment on Full Outer Join query; and

the persistence belief of a response to a pre-assessment on Inner Join query; and then

the persistence belief of failed(N) pre-assessments on Full Outer Join and Inner Join.

In Table 3 is the data and the outcomes of either passed(N) or failed(N) binary states [1, 0] for each leafnode
N and the timespent on each leafnode N pre-assessment task. Recall, Use the "Insert Citation" button to add

Design and analysis of a multi-agent e-learning system using prometheus… (Kennedy E. Ehimwenma)
20 ISSN: 2252-8938

citations to this document.

Use the "Insert Citation" button to add citations to this document. In [33] earlier that we gave the set of
leafnodes N as
N = {union, unionAll, selfJoin, fullOuterJoin, innerJoin, updateSelect, updateWhere,
deleteSelect, deleteWhere, insertSelect, insertValue, selectOrderBy, selectDistinct,
selectWhere, selectAll}.
Now, for the purpose of statistical plots and presentation, let N be encoded a set of positive integers such that
we have N = {1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15}
where N = 1,2, 3, …, 15 are nominal values. Table 3 presents the encoded leafnode N information and the
total number of attempted pre-skill tests. In Table 4 are the passed(N) or failed(N) per the time spent on each
leafnode N, respectively. From Table 3, it is observed that the changes introduced has caused some
significant changes in the results as against our previous findings.

Table 3. Leafnode encoding and number of Passed(N) and Failed(N)

Leafnodes N Leafnode Encoding No. of Passed(N) Pre-assessment No. of Failed(N) Pre-assessment
Union [U] 1 - -
unionAll [UA] 2 - -
selfJoin [SJ] 3 11 2
fullOuterJoin [FOJ] 4 6 7
innerJoin [IJ] 5 4 9
UpdateSelect [US] 6 1 15
updateWhere [DW] 7 16 0
deleteSelect [DS] 8 5 16
deleteWhere [DW] 9 20 1
insertSelect [IS] 10 3 17
insertInto [II] 11 16 4
selectOrderBy [SOB] 12 11 3
selectDistinct [SD] 13 14 1
selectWhere SW] 14 14 0
selectAll [SA] 15 14 0

Visualization of the data is presented in Figures 15-17. The data was plotted using 80% training and 20%
test. Figure 15 shows the scatter plot of the timespent against the leafnode N encoded as integer values with
the display of the respective leafnode N per timespent. In Figure 16 is the scatter plot of linear regression
model. From the plot, the linear model predicts that there would be an increase passed(N) ≡ 1 binary state as
the timespent on pre-assessment tasks decreases. Invariably in the plot, there is a correlation between increase
in passed pre-assessments and continuous decrease in time spent; which implies increase in the
recommendation of chosen desired topics. Figure 17 is a plot of the logistic regression model. Like the linear
regression model, the model also predict increase in passed(N) pre-assessments. That is, in future more
students are likely to pass their pre-assessments in the domain of SQL, if and only if, the two varied
approaches introduced here are kept and adopted.

Table 4. Boolean classification [1, 0] and time spent on each pre-assessment task (continue)
Leafnode
Boolean Classification [1, 0] vs. Timespent (mm:ss)
Encoding
1 Nil | Nil
2 Nil | Nil
[1] | (12:09), [1] | (09:11), [1] | (05:33), [1] | (10.01), [1] | (04:21), [1] | (11:12), [1] | (05:21), [1] | (08: 01), [1] | (04: 07),
3
[1] | (07:25), [1] | (08: 45), [0] | (10:12), [0] | (07:13)
[1] | (05:16), [1] | (13:02), [1] | (10:22), [1] | (06: 56), [1] | (11:34), [1] | (15:08), [0] | (9:19), [0] | (5:33), [0] | (16:48), [0] |
4
(17:59), [0] | (06:41), [0] | (05: 00), [0] | (11:54)
[1] | (05: 55), [1] | (04:35), [1] | (16: 24), [1] | (07: 31); [1] | (02:47), [1] | (06:57), [0] | (09:35), [0] | (09:12), [0] | (11:43),
5
[0] | (05:13), [0] | (11:48), [0] | (13:10), [0] | (14: 19)
[1] | (19:00), [0] | (20:03), [0] | (13: 44), [0] | (07: 11), [0] | (15:17), [0] | (15: 08), [0] | (03:51), [0] | (08: 10), [0] | (02: 14),
6
[0] | (01: 46), [0] | (15: 16), [0] | (18:05), [0] | (11: 10), [0] | (03: 49), [0] | (14:10), [0] | (09:43)
[1] | (01:23), [1] | (01: 58), [1] | (04: 11), [1] | (11:29), [1] | (03:14), [1] | (15:10), [1] | (11.21), [1] | (08:41), [1] | (11:03),
7
[1] | (05:51), [1] | (15: 09), [1] | (04: 17), [1] | (04:16), [1] | (01: 44), [1] | (03:17), [1] | (11: 04)
[1] | (10: 26), [1] | (12:05), [1] | (13: 02), [1] | (17: 33); [1] | (12:24), [0] | (11:15), [0] | (03: 45), [0] | (07: 30), [0] | (11:19),
8 [0] | (05: 18), [0] | (03:55), [0] | (18: 00), [0] | (03: 44), [0] | (21: 40), [0] | (07: 25), [0] | (11:37), [0] | (19: 16), [0] | (02:
41), [0] | (14:12), [0] | (08:13), [0] | (04: 58)

Int J Artif Intell, Vol. 10, No. 1, March 2021: 9 – 23

Int J Artif Intell ISSN: 2252-8938 21

Table 4. Boolean classification [1, 0] and time spent on each pre-assessment task
Leafnode
Boolean Classification [1, 0] vs. Timespent (mm:ss)
Encoding
[1] | (02:22), [1] | (10: 15), [1] | (06: 08), [1] | (06:11), [1] | (05:40), [1] | (02:15), [1] | (01.58), [1] | (02:12), [1] | (17:22),
9 [1] | (09:21), [1] | (08: 39), [1] | (07: 47), [1] | (07:15), [1] | (01: 16), [1] | (12:18), [1] | (01: 54),
[1] | (15:15), [1] | (07: 11), [1] | (11:18), [1] | (11: 54), [0] | (01:32)
[1] | (02:22), [1] | (01:23), [1] | (07: 36), [0] | (01:55), [0] | (01: 57), [0] | (04:10), [0] | (11:31), [0] | (02:20), [0] | (14:25),
10 [0] | (03: 00), [0] | (03: 58), [0] | (08:14), [0] | (06: 47), [0] | (12:37), [0] | (05: 21), [0] | (12:17), [0] | (11:12), [0] | (04:11),
[0] | (07:15), [0] | (08:18)
[1] | (00:59), [1] | (06: 50), [1] | (02: 01), [1] | (01:29), [1] | (01:45), [1] | (03:04), [1] | (03:22), [1] | (04:01), [1] | (07:23),
11 [1] | (05:11), [1] | (02:48), [1] | (04:07), [1] | (04:10), [1] | (09: 10), [1] | (05:31), [1] | (01:23), [0] | (02:59), [0] | (02:21),
[0] | (05: 26), [0] | (13:05)
[1] | (03: 56), [1] | (04:00), [1] | (00:53), [1] | (07:34); [1] | (01:19), [1] | (03:12), [1] | (06:22), [1] | (04:31), [1] | (05: 12),
12
[1] | (07:04), [1] | (07:17), [0] | (01: 40), [0] | (04:51), [0] | (06:33)
[1] | (01:01), [1] | (01: 51), [1] | (02: 28), [1] | (03:12), [1] | (03:41), [1] | (02:35), [1] | (07:48), [1] | (03:27), [1] | (07:16),
13
[1] | (04:43), [1] | (01: 59), [1] | (02: 55), [1] | (04: 17), [1] | (03: 26), [0] | (02: 38)
[1] | (03:21), [1] | (11: 14), [1] | (00: 56), [1] | (05:21), [1] | (04:11), [1] | (02:22), [1] | (00:50), [1] | (02:32), [1] | (00:26),
14
[1] | (03:39), [1] | (02: 01), [1] | (04: 21), [1] | (04: 15), [1] | (04: 44)
[1] | (00:37), [1] | (02: 55), [1] | (00: 28), [1] | (01:58), [1] | (03:01), [1] | (01:45), [1] | (01:13), [1] | (01:42), [1] | (05:29),
15
[1] | (01:17), [1] | (01: 30), [1] | (04: 27), [1] | (00:23), [1] | (00: 34)

Figure 15. Scatter plot of timespent per leafnode N Figure 16. Linear regression plot of timespent to
boolean classification

Figure 17. Logistic regression plot on timespent to boolean classification

5. CONCLUSIONS
This paper has presented a detailed analysis and design of a formative elearning multi-agent pre-
assessment system using the Prometheus methodology design tool and implementation with Jason
(Agentspeak) language. Detailed description and functions of the agent has been presented using different
diagrams of the respective agents and their roles. The paper covered all design activities including issues that
evolved during implementation and the solution strategy that was adopted; then to evaluation, data collection

Design and analysis of a multi-agent e-learning system using prometheus… (Kennedy E. Ehimwenma)
22 ISSN: 2252-8938

and analysis of data. The design activity also covered percept observation by the interface agent, to inter-
agent communication, decision making strategy, classification of user skills and recommendation of materials
for students’ study. The content of the system is databases/SQL: a subject that has been asserted to have pose
difficulty to students. This project was designed to identify the gaps between what a student wants to learn
and what the student has already learned. The two conditions introduced during the pre-skill tests in this
paper have shown that changes in certain factors can change the narratives of the difficulty faced by students
in SQL programming. Further work is the formalization of agent rules using a formalized language in the
pre-assessment and recommendation strategy.

REFERENCES
[1] UML, “Introduction to OMG's Unified Modeling Language,” 5 Feb 2020. [Online]. Available:http://www.omg.org/
gettingstarted/what_is_uml.htm.
[2] S. Stoyanov, I. Ganchev, I. Popchev and M. O’Droma, “An Approach for the Development of InfoStation-Based
eLearning Architectures,” Compt. Rend. Acad. Bulg. Sci, vol. 61, pp. 1189-1198, 2008.
[3] RMIT, “Agents Group,” RMIT University, Australia, 2012. [Online]. Available:
https://sites.google.com/site/rmitagents/software/prometheusPDT/tutorials. [Accessed 5 Apr 2020].
[4] R. H. Bordini, J. F. Hübner and M. Wooldridge, Programming s in AgentSpeak using jason, John Wiley & Sons, 2007.
[5] C. Rouveirol and V. Ventos, “Towards learning in CARIN-ALN,” in In International Conference on Inductive Logic
Programming, Berlin, Heidelberg, 2000.
[6] S. Konstantopoulos and A. Charalambidis, “Formulating description logic learning as an inductive logic programming
task.,” in In International Conference on Fuzzy Systems , 2010.
[7] J. Chen and J. Li, “Globally fuzzy leader-follower consensus of mixed-order nonlinear s with partially unknown
direction control,” Information Sciences,, vol. 523, pp. 184-196, 2020.
[8] W. A. Munassar and A. F. Ali, “Semantic Web Technology and Ontology for E-Learning Environment,” Egyptian
Computer Science Journal, vol. 43, no. 2, pp. 88-100, 2019.
[9] A. S. Aziz, S. A. Taie and R. A. El-Khoribi, “The Relation between the Learner Characteristics and Adaptation
Techniques in the Adaptive E-Learning Systems.,” in International Conference on Innovative Trends in
Communication and Computer Engineering (ITCE), 2020.
[10] A. Trifa, A. Hedhili and W. L. Chaari, “Knowledge tracing with an intelligent agent, in an e-learning platform,”
Education and Information Technologies, vol. 24, no. 1, pp. 711-741, 2019.
[11] U. C. Apoki, S. Ennouamani, H. K. M. Al-Chalabi and G. C. Crisan, “A Model of a Weighted Agent System for
Personalised E-Learning Curriculum,” in International Conference on Modelling and Development of Intelligent
Systems, 2019.
[12] F. El Hajj, A. El Hajj and R. A. Chehade, “Vulnerability detector for a secured E-learning environment,” in Sixth
International Conference on Digital Information Processing and Communications (ICDIPC), Beirut, 2016.
[13] C. Anghel and I. Salomie, “JADE Based solutions for knowledge assessment in eLearning Environments,” EXP-in
search of innovation (Special Issue on JADE), 2003.
[14] N. Stancheva, I. Popchev, A. Stoyanova-Doycheva and S. Stoyanov, “Automatic generation of test questions by
software agents using ontologies,” in In 8th International Conference on Intelligent Systems (IS), 2016.
[15] V. M. F. Ferreira, J. C. C. Carvalho, R. M. E. M. da Costa and V. M. B. Werneck, “Developing an educational
medical game using AgilePASSI multi-agent methodology,” in In 28th International Symposium on Computer-Based
Medical Systems (CBMS),, 2015.
[16] X. Lan, L. Liu and Y. Wang, “ADP-Based Intelligent Decentralized Control for s Moving in Obstacle Environment.,”
IEEE Acess, vol. 7, pp. 59624-59630, 2019.
[17] G. Al-Hudhud, “Designing e-Coordinator for improved teams collaboration in graduation projects,” Computers in
Human Behavior, no. 15, pp. 640-644, 2015.
[18] M. Wooldridge, N. Jennings and D. Kinny, “The Gaia methodology for agent-oriented analysis and design,”
Autonomous Agents and s, vol. 3, no. 3, pp. 285-312, 2000.
[19] L. Cernuzzi and F. Zambonelli, “Gaia4E: A Tool Supporting the Design of MAS using Gaia,” in In ICEIS (4), 2009.
[20] N. R. Jennings and M. J. Wooldridge, “Applications of intelligent agents,” 1998.
[21] T. I. Zhang, E. Kendall and H. Jiang, “An agent-oriented software engineering methodology with application of
information gathering systems for LCC,” in In for LLC, Procs AOIS-2002., 2002.
[22] P. Bresciani, A. Perini, P. Giorgini, F. Giunchiglia and J. Mylopoulos, “Tropos: An agent-oriented software
development methodology,” Autonomous Agents and s, pp. 8(3), 203-236, 2004.
[23] M. Morandini, D. C. Nguyen, L. Penserini, A. Perini and A. Susi, “Tropos Modeling, Code Generation and Testing
with the Taom4E Tool,” in In iStar, 2011.
[24] S. A. DeLoach, “Analysis and Design using MaSE and agentTool,” Air force inst of tech wright-patterson afb oh
school of engineering and management., 2001.
[25] M. Cossentino and C. Potts, “A CASE tool supported methodology for the design of s,” in In International Conference
on Software Engineering Research and Practice (SERP'02), 2002.
[26] M. Cossentino, “From requirements to code with the PASSI methodology,” Agent-oriented methodologies, 3690, pp.
79-106, 2005.
[27] L. Padgham and M. Winikoff, “Developing intelligent agent systems: A practical guide,” 2004.

Int J Artif Intell, Vol. 10, No. 1, March 2021: 9 – 23

Int J Artif Intell ISSN: 2252-8938 23

[28] S. J. Juneidi and G. A. Vouros, “Agent role locking (ARL): theory for multi agent system with e-learning case study.,”
in IADIS AC, 2005.
[29] L. Padgham, J. Thangarajah and M. Winikoff, “Prometheus Design Tool,” in In AAAI, 2008.
[30] Z. Zhang, J. Thangarajah and L. Padgham, “Automated unit testing intelligent agents in PDT,” in In Proceedings of
the 7th international joint conference on Autonomous agents and multiagent systems: demo papers, 2008.
[31] N. Manouselis, H. Drachsler, R. Vuorikari, H. Hummel and R. Koper, “Recommender systems in technology
enhanced learning.,” Recommender systems handbook, pp. 387-415, 2011.
[32] K. E. Ehimwenma, A multi-agent approach to adaptive learning using a structured ontology classification system,
Sheffield UK: Doctoral thesis, Sheffield Hallam University, 2017.
[33] “AUML-2 & Interaction Diagram Tool,” [Online]. Available: http://waitaki.otago.ac.nz/~michael/auml/.
[34] R. H. Bordini, J. F. Hübner and D. M. Tralamazza, “Using Jason to implement a team of gold miners,” in In
International Workshop on Computational Logic in s, Berlin Heidelberg, 2006.
[35] K. Ehimwenma, M. Beer and P. Crowther, “Pre-assessment and learning recommendation mechanism for a Multi-
agent System,” in In 14th International Conference on Advanced Learning Technologies (ICALT) , Sheffield UK,
2014.
[36] K. E. Ehimwenma, P. Crowther and M. Beer, “Formalizing logic based rules for skills classification and
recommendation of learning materials,” International Journal Information Technology and Computer Science
(IJITCS),, vol. 10, no. 9, pp. 1-12, 2018.
[37] K. E. Ehimwenma, M. Beer and P. Crowther, “Computational Estimate Visualisation and Evaluation of Agent
Classified Rules Learning System,” International Journal of Emerging Technologies in Learning (IJET), vol. 11, no.
1, pp. 38-47, 2016.
[38] A. Dennis, B. H. Wixom and D. Tegarden, Systems Analysis and Design with UML, Wiley, 2009.
[39] J. C. Prior, “Online assessment of SQL query formulation skills,” In Proceedings of the fifth Australasian conference
on Computing education, Australian Computer Society, Inc, vol. 20, pp. 247-256, 2004.
[40] K. E. Ehimwenma, M. Beer and P. Crowther, “Student Modelling and Classification Rules Learning for Educational
Resource Prediction in a Multiagent System,” in 7th Computer Science and Electronic Engineering Conference
(CEEC2015), 2015.

BIOGRAPHIES OF AUTHORS

Kennedy E. Ehimwenma has his PhD in s, knowledge representation and rule-based logic in
the area of eLearning application development at the Sheffield Hallam University, United
Kingdom. His research interests include intelligent agent learning, semantic ontology, rule-based
logic, and decision support systems. Dr. Ehimwenma is a lecturer in the Department of
Computer Science, Wenzhou-Kean University. Email: [email protected],
[email protected]. ORCID: https://orcid.org/0000-0002-7616-9342

Sujatha krishnamoorthy is an Assistant professor of Computer Science at Wenzhou-kean

University. Dr. Krishnamoorthy received her Ph. D in Information and Communication
Technology, and her area of research covers digital image processing with image fusion,
machine learning and computer vision. Email: [email protected], [email protected]
[01:28, 04/05/2020] Dr. Sujatha: orcid:0000-0002-0122-6357

Design and analysis of a multi-agent e-learning system using prometheus… (Kennedy E. Ehimwenma)