Adaptive Multiagent System for Learning Gap Identification

Through Semantic Communication and Classified Rules Learning

Kennedy E. Ehimwenma, Martin Beer and Paul Crowther

Communication & Computing Research Centre, Sheffield Hallam University, Sheffield, U.K.

Keywords: Multiagent, Classification, Semantics, Learning, Pre-assessment, Knowledge Representation, Prerequisite,

Jason Agentspeak.

Abstract: Work on intelligent systems application for learning, teaching and assessment (LTA) uses different

strategies and parameters to recommend learning and measure learning outcome. In this paper, we show

how agents can identify gaps in human learning, then the use of a set of parameters which includes

desired_concept, passed and failed predicate attributes of students in the construction of an array of

classified production rules which in-turn make prediction for multipath learning after pre-assessment in a

multiagent system. The context in which this system is developed is structured query language (SQL)

domain with concepts being represented in a hierarchical structure where a lower concept is a prerequisite to

its higher concept.

1 INTRODUCTION

Different set of parameters or approaches have been

used by researchers e.g. Chakraborty, Roy & Basu

(2010), Mills & Dalgarno (2007) and Matsuda et al.

(n.d.) to model intelligent systems for human

learning. However, to the best of our knowledge no

system has been modelled to: 1) diagnose the gap(s)

in students learning using the set of modelling

parameters (desired_concept, passed and failed

predicates) used in this work, 2) use of predicate or

description logic (DL) semantic literals that are

understandable between agents for developing

intelligent tutoring system (ITS), and 3) modelling

production rules for learning material prediction.

1.1 Objectives of Research

 To use some set of parameters that comprises

of desired_Concept, Passed, and Failed

attributes in diagnosing knowledge gaps in

students;

 To use symbolic representation and speech acts

performatives (i.e. communicative acts) in

supporting the development of an intelligent

multiagent based Pre-assessment System; and

 To use production-rule multiple classification

and learning technique by agents in the

development of ITS on the platform of Jason

API.

1.2 Research Problem

How can intelligent agent system identify gaps and

guide students so that they are fully prepared for the

next stage in their learning?

1.3 Related Work

Agents, machine learning and fuzzy logic

approaches have been used in the development of

computer based learning systems. For example,

SimStudent (Matsuda et al., n.d.) was developed

using agent and machine learning for modelling the

system. Chakraborty, Roy & Basu (2010) engaged

the use of two-parameter attributes, namely:

comprehensive ability and problem solving

skill―<C, P>―to develop a student model using a

combination of multiagent system (MAS) and

machine learning. In Mills & Dalgarno (2007), a

combination of machine learning technique and

MAS were also used to provide hints to students on

a current learning goal and to make performance

prediction. In the foregoing analysis of agents and

machine learning approaches to LTA technology,

SimStudent (Matsuda et al., n.d.) was used only for

tutoring, Chakraborty, Roy & Basu (2010) was used

for formative measurement, and Mills & Dalgarno

Ehimwenma K., Beer M. and Crowther P..

Adaptive Multiagent System for Learning Gap Identiﬁcation Through Semantic Communication and Classiﬁed Rules Learning.

 2015 SCITEPRESS (Science and Technology Publications, Lda.)

(2007) for making performance prediction.

However, none of these have mentioned the same set

of parameters, or the use of DL programming tool

nor multiple classification ripple down rule:

MCRDR (Kang et al., 1995) inference method

approach in the development of intelligent learning

and tutoring system as applied in this research.

2 STATE OF THE ART: THE

PRE-ASSESSMENT SYSTEM

2.1 Agents & Multiagent System

An agent is a computer system that is situated in

some environment, and capable of autonomous

action in this environment in order to meet its design

objectives (Wooldridge, 2009). A group of

distributed agents with a collective goal constitutes a

multiagent system. The pre-assessment system is a

multiagent based system (Fig. 1) implemented in

Jason AgentSpeak (Bordini et al., 2007). This

system has five agents which are described as

follows:

2.1.1 Agent agInterface

Observes the dynamic user inputs on the artifact.

2.1.2 Agent agModelling

This is referred to as the classifier. It learns and

classifies the attributes received from the agent

agSupport.

2.1.3 Agent Student Model

This is the agent that constructs and keep track of all

the student activity that is communicated from the

agent agSupport. The Student Model agent uses

Jason TextPersistentBB class. The TextPersistentBB

is a persistent text file that captures all the activity or

learning history of a student which are his: desired

concept, pre-assessment questions, and correct

and/or incorrect answers to questions.

2.1.4 Agent agSupport

This is the teacher in terms of machine learning. It

pre-assesses the student based on his desired state

concept received from the agent agInterface. The

agent also communicate the outcome of pre-

assessment to the agent agModelling and agModel,

respectively, and connects the MAS to the MySQL

database engine for result-set queries.

2.1.5 Agent agMaterial

This is the ontology agent that has all ontological

relations initialised in its BB including the URL data

value of any SQL concept. This agent outputs the

appropriate URL learning material after it is

communicated by the classifier―agent agModelling.

Student

InterfaceAgent



(Classifier)

Model lingAgent



(Pre‐assessment)





Support



Agent



(Ontology)

MaterialAgent



StudentModel



Agent

MySQL



Figure 1: Multiagent System Architecture and Interaction.

2.2 Knowledge Representation

The domain content of the Pre-assessment System is

SQL. This has been represented as an ontology of

class and subclass concepts. To logically establish

the relationships that exist between the concepts of

the ontology, the SQL ontology was modelled,

verified and validated with Protégé 4.3 as a

structured hierarchy of learning content, and then

initialised as internal knowledge model in the agent

agMaterial beliefbase (BB). The agents in the MAS

share, understand and communicate knowledge

about these concepts as messages effectively. From

the ontology, the MAS can:

 test (using a test goal) whether a desired topic

entered by a student exists within the

ontology after it is perceived by the agent

agInterface with a reply confirmation

expected for this request;

 trigger event to return recommended URL links

(data values of leaf-nodes) to students after

the agent agModelling classification of that

student for appropriate learning from its array

of predicated production rules and plans.

The following is a snippet of the knowledge

representation (KR) of the SQL ontology with every

relation semantically annotated as [o(sql)].

Class-to-class relations uses the hasPrerequisite

predicate while the class-to-subclass relations uses

the has_KB relation and has two leaf-nodes. The

CSEDU2015-DoctoralConsortium

subclass to data value relation uses the hasContent

predicate.

hasPrerequisite(union, join)[o(sql)].

has_KB(join, outer_join)[o(sql)].

has_KB(join,inner_join)[o(sql)].

hasPrerequisite(join, update)[o(sql)].

has_KB(update,update_select)[o(sql)].

has_KB(update,update_where)[o(sql)].

hasPrerequisite(upadate, delete)[o(sql)].

has_KB(delete, delete_select)[o(sql)].

has_KB(delete, delete_where)[o(sql)].

hasPrerequisite(delete, insert)[o(sql)].

has_KB(insert, insert_select)[o(sql)].

has_KB(insert, insert_value)[o(sql)].

hasPrerequisite(insert, select)[o(sql)].

has_KB(select, select_where)[o(sql)].

has_KB(select, select_all)[o(sql)].

hasPrerequisite(select, select)[o(sql)].

2.3 Student Model

To identify gaps in students’ learning, the model

agent (shown as agent student in Fig. 2) is modelled

to keep four parameter-information persistently

about a given student in its BB. In a tuple, we

represent this information as: M = <D, P, F, V>

where

M: is the model

D: is a set of desired concepts i.e. desired state

P: is a set of passed pre-assessment i.e. current state gains

F: is a set of failed pre-assessment i.e. current state gaps

V: is the set of SQL query statements.

Parameters <D, P, F> which are also simultaneously

communicated by the agent agSupport to the agent

agModelling are gathered, learned as pre-conditions

within which the appropriate plan is selected to

classify a student and make prediction for his



Figure 2: A snapshot of the model agent “student”

Persistent BB that the human tutor can access and assess

to identify skills gain or gap 24/7 including date and time

stamp of each activity.

learning material. The rule in a Jason plan format for

this classification is given as:

+recommend_material : set_of_profile_parameters

<- recommended_material.

After the agent agSupport pre-assesses a student, it

communicates the student attributes to the agent

agModelling. On receipt of these attributes, the

agModelling classifies and predict some learning

path for the student depending on a desired concept,

passed or failed set of parameters received.

3 METHODOLOGY

3.1 Systems Development

The systems development approach to the pre-

assessment system design is the Agent-Oriented

Analysis and Design (AOAD)―agent software

engineering cycle that involves: systems

specification, analysis, design and implementation.

The AOAD is agent development methodology that

is based on goals, plans and belief. In this work, we

have closely followed the Prometheus methodology

for developing intelligent agent systems (Padgham

and Winikoff, 2004). As a set of guidelines, the

Prometheus methodology involves three major agent

software development steps, given below:

3.1.1 System Specification

This is the phase that presents a scenario of the

problem, identification of goals and basic

functionalities of the system along with inputs

(percepts) and outputs (actions).

3.1.2 Architectural Design

This is the phase where the number and type of

agents are decided. It also consist of 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.

3.1.3 Detailed Design

This phase is focused on the description of

responsibility and development of the internal

structure of each agent, and how they will achieve

their task within the system.

3.2 Goal Specification

As natural construct, goals are central to the

AdaptiveMultiagentSystemforLearningGapIdentificationThroughSemanticCommunicationandClassifiedRules

Learning

functioning of intelligent agent that are going to

realise a functioning system (Padgham and

Winikoff, 2004). The use of goals or subgoals

breakdown a scenario into units of achievable design

steps which map details into later design and

implementation. The following outlines the entire

goal of the MAS that split up into responsibilities for

the various agents:

 Receive Concept

 Fetch subConcept (pre-requisite) quiz

 Test student

 Receive answer from student

 Analyse and pre-assess answer

 Feedback to student

 Update model agent KB

 Classify the student

 Fetch Concept or subConcept materials

 Tutor the student

3.3 Multiple Classification Learning

As the classifier, the agent agModelling learns every

attribute of the parameters it receives from the agent

agSupport during the course of pre-assessment.

Below is an exemplary code in Jason AgentSpeak

from the agent agModelling plan library for a pre-

assessment on the INSERT prerequisite if a

DELETE is received as the desired concept:

/* Prediction rules for DELETE concept */

@d1

+!recommendMaterial[source(agSupport)] :

desired_Concept("DELETE")[source(agSupport)]

& passed("The student has passed the

INSERT with SELECT question.")

& passed("The student has passed the

INSERT with VALUE question.")

<- .send(agMaterial, achieve,

hasPrerequisite(delete, insert)).

@d2

+!recommendMaterial[source(agSupport)] :

desired_Concept("DELETE")[source(agSupport)]

& passed("The student has passed the

INSERT with SELECT question.")

& failed("The student has NOT passed

the INSERT with VALUE question.")

<- .send(agMaterial, achieve, has_KB(insert,

insert_value)).

@d3

+!recommendMaterial[source(agSupport)] :

desired_Concept("DELETE")[source(agSupport)]

& failed("The student has NOT passed

the INSERT with SELECT question.")

& passed("The student has passed the

INSERT with VALUE question.")

<-.send(agMaterial, achieve, has_KB(insert,

insert_select)).

@d4

+!recommendMaterial[source(agSupport)] :

desired_Concept("DELETE")[source(agSupport)]

& failed("The student has NOT passed

the INSERT with SELECT question.")

& failed("The student has NOT passed

the INSERT with VALUE question.")

<-.send(agMaterial, achieve,

hasPrerequisite(insert, select)).

The multiple classification code classifies a student

for learning material into one of four categories for

any given concept e.g. the DELETE. In the codes

the attributes of the students which forms the

production-rules (otherwise known as the context in

Jason agentSpeak) or pre-conditions must be true

and satisfied before classification can be completed.

Recall that the set of parameters that is devised to

construct this multiagent based Pre-assessment

System is given in the turple M = <D, P, F, V>.

Logically, based on the Passed and Failed two-state

predicate attributes of a student, if a set of attributes

are all <P> (e.g. label @d1) then we say the student

has positive ability, but if all <F> (e.g. label @d4)

we say the student has negative ability. But if the set

of attribute is a mix of <P> and <F> (e.g. label @d2

and @d3) then we say it is partial ability.

3.4 Agent Learning Hypothesis

In this production rules classification learning, let C

be the number of prerequisite concept(s) to a desired

concept D, T a binary-state value for student pre-

assessment outcome and N the equal number of leaf-

nodes across each parent node, then the total number

of classified production rules R for a given ontology

tree is determined by:

R = CT

+ 1

where

C ϵ {0, 1, 2, ..., k

-1

, k}

T = 2, for a pass or fail state

N ϵ {1, 2, 3, ..., k

-1

, k}

For any SQL rules set that would need to be added

to the array of production rules, the agent

agModelling would increment the number of

classified rules for a given concept with:

Rꞌ = R + T

;

where C = 0, 1, 2,…, k in

R = CT

+ 1

and conversely decrements by removing rules for a

concept that is no longer needed with:

Rꞌ = R - T

;

where C ≠ 0 in

R = CT

+ 1.

From

each learning algorithm, the number of rules to

CSEDU2015-DoctoralConsortium

be added or removed is determined by the number of

leaf-nodes T

in the ontology. The code snippet in

Section 3.3 gives a picture of how the classified

rules makes prediction for learning. Since T

= 2

then the number

of classified rules equals 4 for each

class concept. In the example for the DELETE

concept received, the agModelling classifies the

student and make prediction for appropriate learning

URL through semantic literal communication to the

agent agMaterial using the tell or achieve

performative.

4 EXPECTED OUTCOME

The main purpose of this system is to ascertain

whether a student is prepared to learn a concept he

so desired to study, and to identify gaps in the

student knowledge. Figure 3 and 4 shows the Pre-

assessment System user interface and a minimized

output console. As students would enter concepts on

the input interface, the MAS outputs a couple of

questions of the immediate lower (or prerequisite)

concept. Depending on the answers received by the

MAS, the student is shown the URL link(s) of the

concept to learn after his/her classification.

Figure 3: Starting the Pre-assessment System. The system

outputs a list from which a SQL concept can be chosen to

be studied.

Figure 4: The user is diagnosed to have a knowledge gap

in INSERT SELECT and INSERT VALUE statement. So

not quite ready for DELETE. Output of INSERT URL

links.

5 STAGE OF THE RESEARCH

The next stage of this work is the evaluation of the

multiagent based Pre-assessment System for: fitness

of purposed, evidence of use, and collation of

feedback data from potential participants and

analysis of the data. The collection of feedback data

shall be through a structured qualitative

questionnaire method, followed by the tutor’s task to

view the student Model agent BB to unveil

successful construction of SQL queries by students

and/or the technical difficulties they are facing

which we have called the gaps between what they

have learnt and what they want to learn in order to

address the research question. Finally write my

Thesis which is potentially titled: Helper Agents for

Learning in Structured Ontology Interoperability

Based System.

6 CONCLUSIONS &

CONTRIBUTION

This paper has demonstrated how a multiagent tool

can be used to design an intelligent learning system,

and how the agents can cooperatively diagnose

gap(s) between a student’s desired knowledge and

his previous knowledge using a devised set of

parameters. As well as the foregoing, its contribution

is the use predicate or description logic semantic

literals that are understandable between agents for

modelling classified production rules and predicting

learning materials. The Expected Outcome is that the

system should be able to pre-assess students,

identify students’ strengths (i.e. gains) for projection

to the next module, identify weaknesses (i.e. gaps) in

order to fill the gaps and prepare them for the next

module; and recommend learning materials.

REFERENCES

Bordini, R. H., Hübner, J. F., & Wooldridge, M. (2007).

Programming multi-agent systems in AgentSpeak

using Jason. John Wiley & Sons, West Sussex,

England.

Chakraborty, S., Roy, D., & Basu, A. (2010).

Development of knowledge based intelligent tutoring

system. Advanced Knowledge Based Systems: Model,

Applications & Research, 1, 74-100.

Kang, B., Compton, P., & Preston, P. (1995, February).

Multiple classification ripple down rules: evaluation

and possibilities. In Proceedings 9th Banff knowledge

AdaptiveMultiagentSystemforLearningGapIdentificationThroughSemanticCommunicationandClassifiedRules

Learning

acquisition for knowledge based systems workshop

(pp. 17-1).

Matsuda, N., Cohen, W. W., Sewall, J., Lacerda, G. &

Koedinger, K. R. (n.d.). SimStudent: Building an

Intelligent Tutoring System by Tutoring a Synthetic

Student. Pp.1-19.

http://www.cs.cmu.edu/~wcohen/postscript/simstudent

-authoring-submitted.pdf (Accessed: January 3

2015).

Mills, C., & Dalgarno, B. (2007). A conceptual model for

game-based intelligent tutoring systems. Proceedings

of the 2007 Australasian Society for Computers in

Learning in Tertiary Education, 692-702.

Padgham, L., & Winikoff, M. (2004). Developing

intelligent agent systems: A practical guide John

Wiley & Sons. West Sussex, England.

Wooldridge, M. (2009). An Introduction to MultiAgent

Systems. John Wiley & Sons Ltd, UK.

CSEDU2015-DoctoralConsortium