An architecture for the dynamic adaptation of an intelligent multi-tutoring system

Abstract

The articulation between machine tutoring and human tutoring remains a productive research within in the context of Intelligent Tutoring Systems (ITS), particularly in the context of e-learning where the dropout rate is high. We explore an innovative approach, the automation of tutoring as it is done in the classroom to respond to the difficulties of the learner. We propose a generic Intelligent Multi-Tutoring System (IMTS) architecture composed of two modules COMES and MAT. The Communication Entry Service (COMES) module manages communications between the IMTS and a Learning Management System (LMS). The module Multi-Agent Tutoring (MAT) is the multi-agent system developed with JADE, which allows the dynamic adaptation of tutoring (Machine, Peer, Teacher) according to the profile of the learner. We offer a configurable system to customize tutoring to the individual needs of each learner. It can be grafted onto any learning platform, making it multidisciplinary and easy to integrate into existing learning environments. The teacher will be able to devote more time to learners which need greater his intervention.The peer will develop human and relational qualities linked to their know-how, transversal skills sought by recruiters. To validate this architecture, we provide an application and results that integrate the elements of the described model. The results of the experiment prove the feasibility and reliability of our approach.

Keywords

Intelligent tutoring system multi-agent system adaptive system markov complex system

1 Introduction

Numerous universities have adapted their educational systems to the new technological environments where pedagogical innovation has become a fundamental challenge. In this context, Intelligent Tutoring Systems (ITS) have grown increasingly popular in the field of online learning. They aim to meet the needs of a new generation of learners. However, the success of such systems is linked to the reaction of their users.

Dropout, abandonment, failure, are all words that indicate that some learners do not complete their training. In open distance learning, dropout is more common than in face-to-face training [1]. The reasons are numerous, but one question arises: What about learners who need the intervention of a human tutor?

Faced with the difficulties commonly encountered (feeling of isolation, difficulty in directing one’s own learning, lack of autonomy or time), we propose a solution to tackle learner abandonment. We believe that the integration of human tutoring in the process of accompanying the learner and the articulation between this human tutoring and the machine tutoring are conditions that are in favor to the learner’s success.

Intelligent tutoring systems have been shown to improve the quality of learning [2, 3]. Several approaches have been implemented to design systems that are customizable and adaptable. In a reactive ITS, the learner’s actions are analyzed with a step by step in a factual way. Each new action triggers a feedback from the system. Two identical actions at different times generate similar reactions from the system. In this classical approach, making the analysis more elaborated and refined consists in considering more situations in which the learner may find his way. Such systems are difficult to scale up and adapt, and consequently they are not very reliable.

In this article, we propose an approach that consists in setting up an intelligent multi-tutoring module that responds to the learner’s difficulties. The idea is to direct the learner to the most suitable tutor for his profile. This module is designed to be grafted on any existing learning platform, which allows it to be with the intelligent tutoring functionalities and thus avoid the learner from quitting.

This paper is organized according to several sections. The first section is devoted to existing work on intelligent tutoring systems. The second section describes our approach for a multi-tutoring articulation. The third section presents the multi-agent architecture of our MAT (Multi-Agent Tutoring). The fourth section describes the implementation of a part of the system. The fifth section presents the results of our study, analyzing the performance of our approach. We end this article with a conclusion that summarizes the contributions of our approach and a short-term perspective.

2 Related work

Intelligent Tutorial Systems (ITS) are defined as computer-based learning frameworks which simulate the way a teacher imparts knowledge to learners through teaching skills [4]. These systems are capable, like a teacher, of accompanying the learner to complete an action and providing the necessary feedback [5].

In addition to the interface with the learner, its architecture composed of an expert module, a pedagogical module and a learner model, echoes the pedagogical triangle knowledge-teacher-learner [6] (Fig. 1). The expert module has an explicit representation of knowledge and reasoning in order to provide the system with the capabilities to solve the problems posed. The learner’s model represents the state of its knowledge and has a diagnostic process using observations (traces of the learner’s actions). Finally, the pedagogical model, thanks to the modeled tutorial strategies, to the learner’s model and to the learning context, dynamically generates its interventions on the choices of contents, of activities and on the exchanges with the learner.

Fig. 1

The components of an ITS.

Didactitians have found that tutoring provides an excellent basis for developing adaptive and interactive systems. They produce individualized and personalized services. In this section, we present a comparative study of some existing ITS.

The main objective of ITS is to adapt learning to the learner’s knowledge and performance. Although ITS helps to facilitate reasoning during learning, there are still challenges and problems to be solved. The study presented in Table 1, reveals that there are two important issues to overcome: multidisciplinarity and multi-tutoring.

Despite the diversity of existing ITS, there is a lack of systems that cover several learning objects. Each ITS is distinguished by a single role (competence, error detection and correction, etc.). There is also a category of ITS that interacts directly with learners by answering their questions.

There is also a lack of human intervention in the learner’s learning process [22]. Existing systems integrate only one type of tutoring (Machine or Peer), which means that the intervention is the same for everyone. Moreover, only one pedagogical strategy, which is implicit, is implemented. The adaptation to the learner is therefore minimalist.

In a classroom setting, the interaction between the teacher and the learner is direct and takes place in several ways. Depending on the learner’s profile, the teacher proposes a solution to remedy his difficulties. The teacher has the ability to know his learners, and then guide them in the best way to remedy their difficulties. The teacher can choose between three solutions:

the learner repeat some or all of the course and its exercises;

choose an advanced learner to help the learner in difficulty;

on the task of explaining the concept that the learner is having difficulty with in another way.

Table 1

Different existing ITS

Domain	Tutoring type
	Machine	Peers
Accounting-Finance	STIMA [7]	ASEPA [8]
Computer science	KERMIT [9]	EPALCP [10]
	CAPIT [11]
Language	SQL TUTOR [12]	COLLECT-UML [14]
	ELMART [13]
	ACTIVE -MATH [15]
Mathematics	BAGHERA [16]	ELEMENTARY [18]
	RADARMATH [17]
Medicine	STOMACH DISEASE [19]
	CIRCUIT ANALYSIS [20]
Physics	THERMO-TUTOR [21]

To overcome these limitations of existing ITS and to better take into account the specificities of learners, other more elaborate software architectures are needed. An architecture that allows for a dynamic and personalized adaptation of the tutoring by also involving the teacher. A solution to reduce the dropout rate of learners would be to offer them an articulation between machine and human tutoring regardless of the learning object.

3 Our approach: Dynamic articulation of tutoring

In the classroom and in case of difficulty, one of the adaptations to the learner’s profile can be made if one of the following two triggers occurs:

the learner is aware of his difficulties and takes the initiative to remedy his problem;

the teacher detects that the learner is in difficulty and decides to intervene.

In both cases, when faced with a difficulty, three possibilities of help are explored:

the learner reviews the course to understand;

the intervention of another advanced learner;

the intervention of the teacher.

Our approach consists in automating this process of difficulty detection and personalized help for the learner. It consists in designing and implementing a system capable of:

detecting that the learner is in difficulty and requires intervention;

identifying and refining the type of intervention suited to the learner’s profile, taking into account their action history.

We propose to integrate the cooperative dimension within the ITS in order to allow all actors (Machine, Peer and Teacher) collaborate to better assist the learners. In face-to-face learning the best choice that adapts to the learner is made at first in an approximate way and over time in a targeted way. The idea that we propose is to articulate the different types of tutoring (Machine, Peer, Teacher) to target the one that best corresponds to the state perceived by the system.

The aim is to offer a multidisciplinary and multi-tutoring solution. A generic module able to accompany the learner in his learning comparable to the face-to-face teaching. We are based on a common point between the different learning objects, those of the solutions adopted to face the difficulties of the learner. The proposed ITS is an independent module interoperable with standard learning platforms. Its role is to extract information from the platform and process it to personalize the tutoring.

It is not a question of totally replacing the teacher’s expertise by the machine, but integrating the teacher into the tutoring triangle (Machine, Peer, Teacher) (Fig. 2). Machine tutoring cannot totally replace the teacher. Peer or teacher tutoring is necessary to meet the expectations that the machine cannot fulfill. Human tutoring offers interactions with the learner, thus avoiding the isolation that leads to abandonment.

Fig. 2

Interaction between Learner, LMS and ITS.

The main objective of this work is to make a tutoring system adaptive, by giving it an intentional character. This means letting the system decide how it behaves with respect to the way it is used. We believe that the feedback of the system should be significantly different when faced with two learners who represent an identical difficulty but have different profiles.

4 Dynamic adaptation of tutoring

A system adapts dynamically during its use when it can change its behavior according to the user’s actions. The system is called upon to regularly make an overall assessment of the learner’s approach, taking into account all his actions on the LMS (Learning Management System).

Based on this assessment, the system can determine the appropriate tutoring for the learner. This dynamic adaptation is made perceptible by more or less important changes in the learner’s actions.

The modeling of such system, whose adaptation is dynamic, requires the definition of its characteristics. It must also ensure an individualized follow-up, which generates a personalized tutoring. This whole process must be independent of the learning object.

4.1 Main features

We define an adaptive ITS as a system that attempts to perceive and approximate a learner’s approach and provide an appropriate response to the learner’s situation. This response is therefore a function of both the learner’s actions and the system’s objective.

To categorize the learner’s approach, the system regularly analyzes all of his or her actions. But in order to allow a real adaptation, we think that the learner’s actions must be analyzed from different points of view.

From this multiple assessment, a dominant point of view generally emerges. The dominant point of view tries to reduce the other points of view and thus propose to the learner the tutoring best adapted to his profile. As the system analyzes the different possible points of view in a competitive way, it decides on its behavior according to this analysis.

It thus has decision-making autonomy, i.e. it determines its own objectives during its operation. For example, it can be perpetually restrictive or incentive-based, depending on its idea of the learner’s profile.

To possess such adaptive behavior, the system must have an internal organization that is in itself adaptive. The system requires a structure that allows for a change of organization. Its autonomy lies in its ability to reorganize itself. Thus, certain internal components must be able to change state, function and type of communication according to the perceived situation. The internal form of the system thus changes by making a dominant point of view emerge, according to the learner’s actions, but which does not totally suppress the other points of view.

From the consideration of its internal organization, the system can determine with relevance the actions, internal and external, that it will perform. The internal actions consist in modifying its own organization. The external actions, the feedback, consist of choosing the dominant pedagogical point of view.

The system therefore adapts its own actions to the situation it perceives. It perceives the learner’s behavior through its actions on the LMS and then adapts its own internal organization, and therefore its own behavior, to this perception. The system acts by adapting to a perceived situation and not by reacting directly to the learner’s actions (Fig. 3).

Fig. 3

Interaction cycle Learner system.

4.2 Intelligent multi-tutoring: Proposed model

We consider the use of a tutoring system in a cyclical way (Fig. 3). From the beginning to the end of the learning path, the learner performs actions that are evaluated by the system. The proposed model is valid for all learning objects and for all levels.

We propose a model composed of three main entities (Fig. 4).

Fig. 4

The three main entities of adaptive tutoring.

The first entity is the learner, the external and main actor who triggers his personalized learning process.

The second entity is the Learning Management System (LMS) whose role is twofold (1) to provide the learning environment (2) to provide, in the form of traces, the learner’s actions and to retrieve feedback from the LMS. The LMS is a Web application developed [23, 24] to allow learners to access courses, complete assignments suggested by the teacher, use educational resources, answer questions and submit their work for evaluation.

The third entity is the Intelligent Multi-Tutoring System (IMTS) module which is responsible for monitoring and personalizing the learner’s learning. The role of this module is to:

recover the learner’s actions in the form of traces from the LMS (learner’s actions).

process these traces to select the type of tutoring best suited to the learner’s profile (feedback).

This process of Learner Actions / System Feedback is continuous. The objective of our approach is to develop an independent IMTS module. The aim is to graft this module to any existing LMS to provide it with an intelligent tutoring. We have two types of interaction: learner/LMS interaction and LMS/IMTS interaction. The role of these interactions is to transfer learner traces and feedback from the system. All of the learner’s traces constitute digital imprints of their own experiences. The feedback corresponds to the intervention (Machine, Peer, Teacher) that is best adapted to the learner’s profile.

5 Architecture of the Intelligent Multi-Tutoring System (IMTS)

The Intelligent Multi-Tutoring System module is composed of two modules: the (COMES) Communication Entry Service and (MAT) Multi-Agent Tutoring module. The MAT module is a complex system [25] for the following reasons:

the situation to represent is dynamic and incremental. The traces that the IMTS module analyzes and processes to provide a result are heterogeneous and uncertain. IMTS must take this evolution into account and adapt its result during its execution;

IMTS is independent of the learning object. This constraint contributes to the complexity of IMTS, since it is necessary to achieve modularity in the design and realization of this module.

The system is composed of two modules COMES (Communication Entry Service) and MAT (Multi-Agent Tutoring) (Fig. 5):

Fig. 5

The two modules of IMTS.

The figure (Fig. 6), using the BPMN (Business Process Model and Notation) formalism [26], provides an overview of the system and also the structure of the rest of the article. Each element of the IMTS is identified and linked to the corresponding sections and subsections: (5. Architecture of the IMTS, 5.1. The COMES module, 5.2. The MAT module, 5.2.1. Detection Organization, 5.2.2. Intervention Organization.)

Fig. 6

BPMN formalization of IMTS.

5.1 The COMES module

The Communication Entry Service (COMES) module manages the communication between the LMS and IMTS. This module is the intermediary between the LMS and the MAT module. Its role is twofold (1) it stores the traces coming from the LMS to send them to the MAT core. (2) It receives the results from the MAT core to transfer them to the LMS. Currently the data formats supported by this interface are XML or Json. In this article, this module is not treated.

5.2 The MAT module

Multi-Agent Tutoring (MAT) module is in charge of processing the continuous flow of data received from the COMES module. It ensures a permanent monitoring of the learner’s status and intervenes if necessary. The rest of this article is mainly dedicated to the MAT module. We propose a multi-agent architecture [27, 28] which allows to make the system adapted according to the described model. The agent paradigm makes it easier for us to take into account the incremental and uncertain nature of data flows and to improve the autonomous learning of learners and the rapid and accurate retrieval of knowledge [29]. This architecture is composed of two agent organizations (Fig. 7).

Fig. 7

Multi-Agent intelligent multi-tutoring system.

The MAT module is made up of two agent organizations that structure their collective work. Agents interact by sending messages. An agent can participate in a dialog either passively or actively. A passive agent accepts and answers questions from other agents. An active agent proposes and sends questions. In a dialogue, agents alternate between active and passive roles, exchanging a series of messages according to precise coordination and negotiation protocols.

Coordination enables agents to manage their commitments in the event that the circumstances in which they were drawn up change. Coordination also defines the conditions under which agents’ commitments can be reviewed, and the actions to be taken. Negotiation occurs when agents interact to make joint decisions, even though they are pursuing different goals. To exchange information and knowledge, agents use ACL (Agent Communication Language), and in particular ACL-FIPA [30].

5.2.1 Detection Organization

During the work session, the detection agents filter the learner’s actions in order to keep only those that correspond to their progress. They factually analyze the learner’s actions to determine if they are normal and if not to indicate the degree of impact.

Each of these agents reviews the learner’s qualified actions. It communicates its status to the other agents in the organization. This organization determines instructions to trigger or not the intervention process. The autonomy of the agents allows the system to have various behaviors. The agents analyze each new action, not in isolation, but in relation to past actions deemed relevant.

The sensing agents are strengthened if the learner’s actions match their expectations. In other words, the learner’s actions reinforce or weaken the agents’ states. The internal communications of this organization, its modifications and transformations will represent the interpreted meaning of the emerging current situation.

Each detection agent has a macro state-automaton structure representing the evolution of its state. This macro-automaton is composed of four states:

the Waiting state: This is the state in which the agent is when he/she has not received any external solicitation;

the state of Awakening: The agent is in this state as soon as an action of the learner has been judged relevant;

the state of Reinforcement: This is the state where the agent evaluates external requests to reinforce itself. It must then communicate with the other agents about the situation perceived by the system;

the Action state: This is the state where the agent is certain that the situation corresponds to what he expected. It will trigger the intervention process.

5.2.2 Intervention organization

This organization is concerned with making a probabilistic decision, comparable to the behavior of a teacher faced with the difficulties of a learner. It is composed of different types of agents (Fig. 8):

"Learning Object" agent: its role is to keep track of the different parameters that define a given learning object and to calculate the weighted average.

"Markovian" agent: its role is to redefine an action plan each time the indicators change and to select the optimal intervention based on Markov chains.

"Machine" agent: its role is to redirect the learner to the part of the course to be reviewed.

"Peer" agent: its role is to select a peer among several possible ones.

"Teacher" agent: its role is to select a teacher among several possible ones.

Fig. 8

Organization of intervention agents.

This agent organization needs two primary elements:

the relevant indicators to represent a learning object;

quantifiable methods to address the perceived situation.

This organization is responsible for replicating the behavior of the face-to-face teacher. We know that models of learning lack rigorous assessment methods. The criteria for observation differs from one teacher to another and from one pedagogy to another. For each teacher, the set of criteria, indicators and their values are defined in an approximate way.

The system will need a set of relevant indicators to identify a learner’s situation. However, it is impossible to establish such a list, given the diversity of learning objects. We propose an open system that can be parameterized by the domain expert. Thus, there is a permanent and collaborative interaction with the teacher. In this way, at any time, the teacher can add or remove indicators and also modify the intervals of the feedback orders.

One of the strengths of our system is the customization of parameters and a personalized analysis method. On the one hand, the pedagogical expert chooses and sets the necessary indicators to represent his learning object. On the other hand, the system is equipped with analysis methods the current situation, which are probabilistic but can be parameterized and customized by the educational expert.

5.2.2.1. Indicators. Since our system is multidisciplinary, the indicators depend on the learning object. The expert chooses, in a first step, the indicators he considers necessary to define a learning object (Fig. 9).

Fig. 9

Choice of appropriate indicators.

An indicator is a vector composed of a set of pairs (parameter, weight) that characterize expected knowledge or skills (see formula 1).

$I_{i} ((p_{1}, w_{1}) . . . (p_{m}, w_{m}))$ (1)

I is the set of indicators, which means that we can have one or more indicators. Each indicator I_i is represented by one or more parameters (p₁, p₂, . . . , p_m) with their respective weights (w₁, w₂, . . . , w_m). So, for a specific indicator I_i, we can have a combination of parameters and weights.

The number of indicators, the number of parameters and the weight of each parameter are defined by the domain expert.

Once the expert has completed this initial setup phase, the system calculates the weighted average for each indicator (see formula 2).

$W_{i} = \frac{\sum_{j = 1}^{m} (p_{ij} * w_{ij})}{\sum_{j = 1}^{m} w_{ij}}$ (2)

W_i is the weighted average of I_i. The formula breaks down as follows:

p_ij is the $j^{th}$ parameter of indicator I_i.

w_ij is the weight associated with the $j^{th}$ parameter of indicator I_i.

The "LearningObject" agent continuously retrieves the learner’s traces, the values of the indicators, to calculate their weighted average. This agent transmits this weighted value to the "Markovian" agent which triggers the intervention process.

5.2.2.2. Action Selection. The system proposes to remedy the learner’s difficulty by offering three types of intervention: Machine (M), Peer (P) or Teacher (T) intervention. These interventions are presented in 6 triplets as an arrangement of 3 among 3, and in an ordered manner, e.g. (M, P, T), (M, T, P), (P, M, T)...

The expert associates a level of certainty with each type of intervention, reflecting its reliability. In addition, he defines confidence intervals [0, f1]... [fi, 100] for all the indicators, corresponding to an evaluated assessment of an intervention. The number of intervals is not limited, but the upper limit is set at 100 (Fig. 10).

Fig. 10

The six possible combinations are their confidence intervals.

For each Indicator, provided by the LMS, a correlation is made between its weighted average and the confidence intervals, to select a single triplet of intervention order:

Examples:

(I₁ -> (M, P, T)).

(I₂ -> (M, T, P)).

(I₃ -> (P, M, T)).

(I_n -> (M,T, P)).

So, for all the indicators, we’ll have, as a finality, a matrix (nbrIndicator, nbrIntervention=3) see (Fig. 11) for each combination of intervention triplet.

Fig. 11

The intervention range with percentage certainty for each indicator.

Each level of the intervention order (3 levels) is associated with a certain coefficient X (X₁, X₂, X₃). These coefficients are chosen manually and approximately by the expert to guide the system in its decision-making. The Markov agent calculates the probability of success of each intervention (AP), which represents its Action Plan.

To select an action plan, an algorithm calculates the number of occurrences of possible actions M, P or T by level of certainty. The variable OccM₁ designates the number of occurrences of action M in level 1, OccP₂ designates, for example, the number of occurrences of action P in level 2.

Fig. 12

Markov decision process.

The percentage of success of the Action Plan is defined by calculating the number of occurrences of an action per level and multiplying it by its certainty coefficient divided by the number of indicators (Fig. 11).

For example, the probability of success of the Machine intervention is obtained by the following formula.

For the Machine Action Plan:

$\begin{matrix} AP (M) = \frac{\sum_{k = 1}^{k = 3} ({OccM}_{k} * X_{k})}{n} \end{matrix}$ (3) To observe the following parameters (Fig. 11):

(OccM₁, OccM₂, OccM₃): represents the number of occurrences of variable M at the level of (1, 2, 3).

(X₁, X₂, X₃): certainty coefficient at level (1, 2, 3).

n: number of indicators.

For the Peer Action Plan: $\begin{matrix} AP (P) = \frac{\sum_{k = 1}^{k = 3} ({OccP}_{k} * X_{k})}{n} \end{matrix}$ (4)

For the Teacher Action Plan:

$\begin{matrix} AP (T) = \frac{\sum_{k = 1}^{k = 3} ({OccT}_{k} * X_{k})}{n} \end{matrix}$ (5)

5.2.2.3. Type of Intervention. The implementation of an action plan is subject to a Markov decision process for the selection of the optimal intervention (Fig. 12).

The "Markovian" agent handles probabilistic (approximate) decision making, comparable to the behavior of a teacher. Each feedback from the system calls for a new action from the learner. This new action reinforces or weakens the current intervention state using a bonus/penalty. Thus, the system becomes adaptable to the learner’s actions. The feedback of the system involves a new action of the learner in a continuous way. To do so, the agent uses a transition probability AP/100, representing the probabilities associated with each state in which it may find itself after performing an action from the current state. In order to optimize its decision according to the targeted objectives.

5.2.2.4. Targeted Intervention. One of the key aspects of our system is targeted intervention. The "Markovian" agent is the brain of this intervention as it makes probabilistic decisions based on the nature and complexity of the learner’s difficulty. This agent uses a Markov model to estimate the probabilities of success of each possible type of intervention (Machine, Peer or Teacher). The Machine, Peer and Teacher agents are designed to interact directly with the learner.

The Markov agent decides which type of intervention is most appropriate. When the machine agent is selected, it analyzes the learner’s difficulty and looks for the appropriate response. This response can be in the form of a text, image, etc. The peer agent is responsible for finding the peer(s) who can best help the learner. When this agent is selected, it uses recommendation algorithms [31] to find a peer with the necessary skills to answer this difficulty. This learner matching process is based on criteria of skill dissimilarity, learning levels, and learning preferences. Once the support learner is identified, the agent sends a notification to this peer to assist the learner. Finally, when the teacher agent is selected, a teacher is notified of the learner’s difficulty and provides individualized and personalized assistance.

Overall, the selection of one of these three agents is based on the nature of the difficulty encountered by the learner and the skill required to solve it. By coordinating these different agents, MAT can provide targeted and personalized assistance that effectively helps learners overcome their difficulties and progress in their learning.

6 Implementation

The implementation version that we describe in this section, focuses on the organization of intervention agents. We have implemented this organization under JADE [32]. JADE is a multi-agent platform written in JAVA. The choice of this framework was motivated by the fact that JADE is very responsive, functional and compliant with FIPA [33] standards. JADE uses the notion of container to designate an organization of agents. In our case, we added two containers (detection and intervention) to the JADE Main Container (Fig. 13).

Fig. 13

JADE implementation of the MAT architecture.

The Main Container is used to deploy and run the agents and containers that register with it at startup. It is responsible for maintaining the list of active agents and making it available to agents who request it. The Main Container therefore plays an important role in the communication between agents, it is the main element of the JADE platform and consists of the following elements:

ACC (Agent Communication Channel): This feature manages the communication between agents in the JADE platform. It allows agents to send messages to each other and to communicate with each other. DF (Directory Facilitator): The DF is a "yellow pages" service for the JADE platform. It allows agents to search and locate other agents registered in the platform.

AMS (Agent Management System): The AMS is responsible for overseeing the registration, authentication and access to agents in the JADE platform. It also manages the use of the system by tracking agent activities and ensuring system security.

JADE follows FIPA standards. A JADE message consists of several parts in addition to the content, for example: the intended recipients, the sender and the message type. It is essential that the message as a whole respects a common format. In JADE, messages strictly adhere to the ACL (Agent Communication Language) standard, which allows several possibilities for content encoding. In our context, messages use performative such as INFORM, whereby one agent gives another some useful information, REQUEST to ask the other to perform an action, QUERY to ask a question and PROPOSE to start negotiating. Performatives for responses include AGREE or REFUSE.

For functional reasons, we have developed an LMS to collect the learner’s traces. As we mentioned in the previous section, the expert’s parameterization of the system conditions its good functioning and its behavior both at startup and during execution. For this purpose, we have developed an interface allowing him to set the parameters of the system before its launch.

The (Fig. 14) represents the launch interface. This interface allows the expert:

to view and act on the different indicators.

to launch the intervention process and therefore the MAS.

Fig. 14

Interface for setting up and launching MAT.

Box 1: The first section is located on the right side of the interface. This section is dedicated to indicators that subject matter experts have included in their learning objects. Thanks to this part, we are able to consult the values of the different parameters and make a precise follow-up of each learner’s learning. In fact, this functionality allows us to measure the evolution of the learner’s acquired skills over time, to identify possible gaps and to propose personalized pedagogical adjustments accordingly.

Box 2: Allows us to display the weighted average of each indicator. This information gives us an overview of the learners’ performance for each indicator evaluated. Indeed, by calculating a weighted average, our system takes into account the relative importance of each indicator calculated and allows the subject matter expert to better understand the strengths and weaknesses of each learner.

Box 3: The third box of the interface, the Action Plan, allows the learning object expert to assign to each indicator a specific combination according to its range of values.

Box 4: The fourth box, called "Action Plan", displays the calculation of the Probability of Action (AP) for each type of intervention.

Box 5: This feature gives us the ability to configure our Markov process to assign bonuses and penalty to each action taken by the agent. With this feature, we can adjust the rewards and penalties given to each action, based on the needs and performance of each learner. This customization of rewards and penalties helps to improve the effectiveness of the intervention according to the subject matter expert’s vision.

The optimal intervention, the most adapted to the learner’s profile, is sent to the learner in the form of a dialog box that is displayed on his LMS (Fig. 15).

Fig. 15

Example of optimal intervention in an LMS.

7 Results

The context in which we conducted our experimentation consists of opposing two groups of learners, each composed of 40 learners coming from the same training center, on the same learning object, a PHP course. The first group used a classical e-learning system (LMS), and the second group used the same LMS and our IMTS.

We initially envisioned using an existing LMS to collect learner traces. However, after a thorough analysis of the functionalities required for our MAT module, we began by studying the output of existing LMSs such as Moodle, Chamilo, Canvas and Ilias [34]. However, due to the lack of available data, we decided to develop our own LMS. This solution enabled us to define and determine how well our system would work.

The main contribution of our system, as we will see at the end of this section, is a perceptible improvement in the quality of learning in terms of information sharing and a decrease in dropout times.

7.1 Course of the experiment

We chose to use three criteria to compare the results of the two groups (control group and experimental group):

the number of requests for help and information sharing: indicates whether learners are willing to help each other and share their knowledge.

the quality of feedback. Feedback refers to feedback from the Machine, Peer or Teacher. The quality of feedback measures the Relevance, Clarity, Accuracy and Objectivity of its feedback.

E-learning dropout rate: measures learner motivation and engagement.

Using only these three criteria for comparison, the results obtained reflect the effectiveness and relevance of our approach. We present below the results obtained.

Fig. 16

Comparison of the number of help requests and the rate of information sharing between the control group and the experimental group.

The results of the experiment showed a significant increase in the number of help requests from learners in the experimental group compared to the control group. We observed a shift from 98 to 120 requests, which represents a 23% increase. Information sharing also increased by 37%.

Fig. 17

Comparison of the average score obtained by the experimental and control groups for the evaluation criteria of feedback.

The results of the quality of feedback showed a significant improvement in the experimental group compared to the control group. To measure this improvement, we developed a questionnaire for learners in both groups (control and experimental). The feedback from teachers and peers in the experimental group obtained an average score of 4.1 on a scale of 1 to 5, while the control group only achieved 3.2, resulting in a 28% increase.

Fig. 18

Comparison of the reduction in dropout rate between the control group and the experimental group based on the time spent on the platform.

The results showed a significant decrease in the dropout rate of online learning in the experimental group compared to the control group. To measure this decrease, we used time spent on the platform as a criterion. Specifically, we found that 32% of control learners dropped out of their training after spending less than 24 hours on the platform, compared to 21% of learners in the experimental group.

7.2 Experiment summary

The comparative study demonstrates that the results obtained from the experimental group are significantly better than those from the control group (see Table 2).

Table 2
Summary of findings from experimental and control group

Criterion Indicator/ Experimental Results

Witness

Request for help 98 120 23%

Sharing of information 43 59 37%

Quality of feedbacks 3.2 4.1 28%

Abandonment rate 32 21 34%

Criterion	Indicator/	Experimental	Results
Request for help	98	120	23%
Sharing of information	43	59	37%
Quality of feedbacks	3.2	4.1	28%
Abandonment rate	32	21	34%

The table shows an increase in the rate of information sharing and requests for help among learners, an improvement in the quality of feedback, and a significant decrease in the dropout rate from online learning. These results are important because they show that our system is effective in improving learning by personalizing the pedagogical strategy and increasing collaboration and mutual aid among learners.

In conclusion, our multidisciplinary MAT offers an adaptive tutoring system, configurable by the expert for each learning object and customizable for the learner. The strength of this system is its behaviorist quality, it behaves differently from one expert to another and from one learner to another.

8 Conclusion & perspective

An ITS is a computerized learning environment designed to help a learner learn a particular area of knowledge. Existing ITS are traditionally structured around a single pedagogical module and developed in a single learning object.

This article is a contribution to addressing the limitations of existing ITS. It proposes a multi-tutoring and multi-object learning module to considerably reduce the failure or drop-out rate. We have designed and developed a system that is essential to the teacher’s activity and places learners at the heart of active learning, offering them tools for inquiry-based learning methods and collaborative workspaces. The learner thus has a single ITS for all learning objects, with tutoring based on articulation between Machine, Peer and Teacher. Another original feature of our system is that it can be grafted onto an existing LMS, providing feedback by dynamically adapting tutoring to the learner’s profile.

Our aim is to provide learners with an additional means of learning, and teachers with a means of monitoring learners more effectively. For the teacher, ITS offers the means to devote more time learners who present difficulties and require support that Machine or Peer tutoring cannot provide. For peers, it enables them to develop transversal skills such as mutual aid and communication.

We are working in the context of the development of complex, dynamic and reactive systems, capable of adapting rapidly and progressively to the changing needs and uses of their users. We propose a multi-agent architecture to reflect the complexity and dynamics of the system.

The promising results of the experiment demonstrate and validate the feasibility and reliability of our approach. We are currently exploring Case-Based Reasoning (CBR) [35] and recommendation algorithms [36] to refine both the detection of difficulties and the intervention of the tutor. We will continue experimenting with larger groups of learners and with different learning objects.

References

Ennaji

, Boukachour

and Machkour

, Collaborative tutoring architecture: A generic case based reasoning multi-agent, International Journal of Information Science and Technology 5(2) (2021), 40–50.

Thapliyal

, Ahuja

N.J.

, Shankar

, Cheng

and Kumar

, A differentiated learning environment in domain model for learning disabled learners, Journal of Computing in Higher Education 34(1) (2022), 60–82.

Chaudhry

M.A.

and Kazim

, Artificial intelligence in education (aied): A high-level academic and industry note 2021, AI and Ethics (2022), 1—9.

Campbell

H.G.

, Dudek

R.A.

and Smith

M.L.

, A heuristic algorithm for the n job, m machine sequencing problem, Management Science 16(10) (1970), B–630.

, Feng

and Liu

, Research on the influence of computer aided intelligent tutoring system on teacher’s selfefficacy, Journal of Intelligent & Fuzzy Systems 35(3) (2019), 2749–2759.

Kenny

and Pahl

, Intelligent and adaptive tutoring for active learning and training environments, Interactive Learning Environments 17(2) (2009), 181–195.

Notargiacomo

and Marin

, Multiagent intelligent tutoring system for financial literacy, International Journal ofWeb- Based Learning and Teaching Technologies (IJWLTT) 17(7) (2021), 1–13.

Malan

and Stegmann

, Accounting students’ experiences of peer assessment: A tool to develop lifelong learning, South African Journal of Accounting Research 32(2-3) (2018), 205–224.

Suraweera

and Mitrovic

, Kermit: A constraint-based tutor for database modeling, in International Conference on Intelligent Tutoring Systems, Springer (2002), 377–387.

10.

Sitthiworachart

and Joy

, Effective peer assessment for learning computer programming, ACM SIGCSE Bulletin 36(3) (2004), 122–126.

11.

Mayo

, Mitrovic

and McKenzie

, Capit: An intelligent tutoring system for capitalisation and punctuation, in Proceedings international workshop on advanced learning technologies. IWALT 2000. Advanced learning technology: Design and development issues, IEEE (2000), 151–154.

12.

Tahir

, Mitrovic

and Sotardi

, Investigating the causal relationships between badges and learning outcomes in sqltutor, Research and Practice in Technology Enhanced Learning 17(1) (2022), 1–23.

13.

Weber

and Brusilovsky

, Elm-art–an interactive and intelligent web-based electronic textbook, International Journal of Artificial Intelligence in Education 26(1) (2016), 72–81.

14.

Baghaei

and Mitrovic

, Collect-uml: supporting individual and collaborative learning of uml class diagrams in a constraint-based intelligent tutoring system, in Knowledge- Based Intelligent Information and Engineering Systems: 9th International Conference, KES Melbourne, Australia, September 14-16, 2005. Proceedings, Part IV 9. Springer, (2005), 458–464.

15.

Melis

and Siekmann

, Activemath: An intelligent tutoring system for mathematics, in International Conference on Artificial Intelligence and Soft Computing, Springer (2004), 91–101.

16.

Trgalova

and Soury-Lavergne

, Diagnostic de conceptions d’eleves dans baghera, un environnement informatique pour l’apprentissage humain, 2003.

17.

, Pian

, Chen

, Meng

and Cao

, Radarmath: An intelligent tutoring system for math education, in Proceedings of the AAAI Conference on Artificial Intelligence 35(18) (2021), 16087–16090.

18.

Herbert

and Bragg

L.A.

, Elementary teachers’ planning for mathematical reasoning through peer learning teams, International Journal for Mathematics Teaching and Learning 22(1) (2021), 24–43.

19.

Almurshidi

S.H.

and Abu-Naser

S.S.

, Stomach disease intelligent tutoring system, 2017.

20.

Butz

B.P.

, Duarte

and Miller

S.M.

, An intelligent tutoring system for circuit analysis, IEEE Transactions on Education 49(2) (2006), 216–223.

21.

Mitrovic

, Williamson

, Bebbington

, Mathews

, Suraweera

, Martin

, Thomson

and Holland

, Thermo-tutor: An intelligent tutoring system for thermodynamics, in 2011 IEEE global engineering education conference (EDUCON). IEEE (2011), 378–385.

22.

Ennaji

, Boukachour

, Machkour

and Salah

M.B.

, Mutual human-computer tutoring using agents, in 6th IEEE Congress on Information Science and Technology (CiSt). IEEE (2021), 279–283.

23.

Bradley

V.M.

, Learning management system (lms) use with online instruction, International Journal of Technology in Education 4(1) (2021), 68–92.

24.

Qendraj

D.H.

, Xhafaj

and Thanasi

, A hybridized approach of pls-sem and fuzzy z-ahp to evaluate the utaut2 model for an lms, Journal of Intelligent & Fuzzy Systems, no. Preprint (2022), 1–13.

25.

Stieler

, Schwinn

, Leder

, Maierhofer

, Kannenberg

and Menges

, Agent-based modeling and simulation in architecture, Automation in Construction 141 (2022), 104426.

26.

von Rosing

, White

, Cummins

and de Man

, Business process model and notation-bpmn, 2015.

27.

Jiang

and Lu

, Learning attentional communication for multi-agent cooperation, Advances in Neural Information Processing Systems 31 (2018).

28.

Cardoso

R.C.

and Ferrando

, A review of agent-based programming for multi-agent systems, Computers 10(2) (2021), 16.

29.

, An artificial intelligence agent technology based web distance education system, Journal of Intelligent & Fuzzy Systems 40(2) (2021), 3289–3299.

30.

Louis

and Martinez

, An operational model for the fipaacl semantics, in International Workshop on Agent Communication, Springer (2005), 1–14.

31.

Khanal

S.S.

, Prasad

, Alsadoon

and Maag

, A systematic review: machine learning based recommendation systems for e-learning, Education and Information Technologies 25 (2020), 2635–2664.

32.

Bellifemine

, Bergenti

, Caire

and Poggi

, Jade–a java agent development framework, In Multi-agent Programming, Springer (2005), 125–147.

33.

Bergenti

, Caire

, Monica

and Poggi

, The first twenty years of agent-based software development with jade, Autonomous Agents and Multi-Agent Systems 34 (2020), 1–19.

34.

Ouariach

, Khaldi

, Mohamed

and Khaldi

, From the choice of a learning management system to the installation of a platform in a server, in Handbook of Research on Scripting, Media Coverage, and Implementation of E-Learning Training in LMS Platforms. IGI Global, (2023), 330–375.

35.

Richter

M.M.

and Weber

R.O.

, Case-based reasoning. Springer, 2016.

36.

, Zeng

, Gillard

and Medo

, Network-based recommendation algorithms: A review, Physica A: Statistical Mechanics and its Applications 452 (2016), 192–208.