Intelligent techniques in e-learning: a literature review

Learners have different motivations, prior knowledge, personalities, emotions, and learning habits, all of which can have an impact on their educational process (Abyaa et al. 2019). As a result, supplying each learner with tailored learning content is no longer an option, but a necessity. Understanding the learners may help teachers design better instruction and materials, but it can also assist learners in becoming more aware of how they learn best.

Learner modeling is the act of gathering and updating data about a learner over time using well-defined procedures. The stages of this procedure are (1) obtaining initial data on the learner’s attributes, (2) model construction, and (3) updating the learner model by watching and tracing the learner’s activities (Abyaa et al. 2019; Vagale and Niedrite 2012).

Literature review, presented in Abyaa et al. (2019) analyzes the works in the area and identifies three main areas of interest regarding learner modeling: the modeling approaches, the modeling techniques, and the learner characteristics being modeled. In addition, we identified another aspect of learner modeling that should be investigated in more detail - used technologies and standardization efforts.

The learner model can be implemented using two different modeling approaches: knowledge-based and behavioral-based (Ahmed et al. 2017). These two approaches are used in e-learning to acquire substantial information about learners needed to create a learner model. Knowledge-based approach gathers learner data through examination, questionnaires, learner interests, and learners’ learning routines. The behavioral-based approach collects learner data only by observing learners’ system activities and processes.

Regarding the question of which learner’s data is being collected, six major categories of modeled characteristics were identified in the literature, depending on the used approach: the learner profile (mostly learner personal information and static data), knowledge (knowledge level, competencies, skills errors, misconceptions), cognitive characteristics (learning and thinking styles, cognitive states, learner’s behavior), social characteristics (interactions, culture, social style), motivation (interests, learning goals, engagement and affect), and personality (Abyaa et al. 2019). These characteristics represent the type of learner data that is being collected and then used for modeling.

Influenced by the modeling approach and the defined goals of e-learning system, five modeling techniques can be identified (Abyaa et al. 2019): predictive modeling techniques (De Morais et al. 2014), clustering and classification techniques (Moubayed et al. 2020), overlay modeling (Ahmadaliev et al. 2019), uncertainty modeling (Al-Shanfari et al. 2017), and ontology-based learner modeling (Akharraz et al. 2018; Rezgui et al. 2014). These techniques are used for modeling learner models using collected data.

Some typical examples of e-learning systems that use a knowledge-based approach for learner modeling are Zebra (Nguyen 2014) (ontology-based learner modeling technique, with modeled characteristic knowledge), AdaptLearn (Alshammari et al. 2015) (uncertainty modeling technique, with cognitive modeled characteristics), and Learning Java (Yau and Hristova 2018) (uncertainty modeling technique, with cognitive modeled characteristics). Some typical examples of e-learning systems that use a behavioral-based approach for learner modeling are DEPTHS (Jeremić et al. 2012) (overlay technique, with modeled characteristic knowledge), INSPIREus (Papanikolaou 2014) (clustering and classification technique, with modeled characteristics knowledge and motivation), and My Math Academy (Thai et al. 2021) (predictive modeling technique, with modeled characteristic knowledge).

Numerous online courses and e-learning platforms have been created independently, frequently at great expense. Furthermore, such content and systems are often, if not always, unprepared to communicate or share data. Systems must be able to read and understand other systems’ data structures to communicate and interoperate, therefore need extensive standardization efforts. Shareable Content Object Reference Model, or SCORM, is a set of technical standards for e-learning software products (Poltrack et al. 2012). Due to current web trends and the SCORM standard inflexibility for formal education, a new strategy and transition to the development of a new standard were needed, and it was necessary to switch to Experience API (xAPI). The xAPI represents a specification that enables different learning technologies to capture data about a person’s or a group’s wide range of learning experiences in a consistent format using the xAPI vocabulary (Karoudis and Magoulas 2018). It provides a learner-centered model for learning data collection and learning process recording (Sun et al. 2020). With this technology, experiences can be collected in the form of statements and stored in a learner record store that syncs learning activities across platforms and devices (Smith et al. 2018; Zapata-Rivera and Petrie 2018). As a result, learning applications could make use of much more complete data to support learners with visualizations and interventions (Neitzel et al. 2017). E-learning systems can more effectively adapt instruction when they have more accurate models of the learner’s prior knowledge or competency (Sottilare et al. 2017).

Data mining is an effective tool to extract meaningful and interesting patterns from the current and historical data stored in data warehouses or data repositories, to be analyzed and predict future trends (Prabha and Shanavas 2015). In the context of educational research, data mining is known as Educational Data Mining (EDM). EDM is defined as the area of scientific inquiry centered around the development of methods for making discoveries within the unique kinds of educational data and using those methods to understand learners better and the environments in which they learn (Baker 2010).

The most common uses of EDM are for supporting learners in course selection, learners’ profiling, finding problems leading to dropout, learners’ targeting, curriculum development, predicting learners’ performance, and as a support for decision-making at learner enrollment (Zorić 2020).

The benefits of EDM are numerous, including enhancing the quality of education, improving current study programs and educational practice, improving teaching, advancing the process of studying, improving learners’ academic performance, reducing learners’ failure rates, increasing course completion percent, and helping educational management to be more efficient and effective (Abu Tair and El-Halees 2012; Kumar and Chadha 2011; Zorić 2020).

The five most commonly used techniques in the educational domain found in the literature are: prediction, clustering, relationship mining, distillation for human judgment, and discovery with models (Bienkowski et al. 2012). Each of these techniques can be used to quantitatively analyze large data sets to find hidden meaning and patterns (Huebner 2013).

Prediction entails developing a model that can infer a specific element of the data (predicted variable) from a set of predictor variables (Bienkowski et al. 2012). This technique is used in e-learning to analyze learner data and predict outcomes. The types of prediction methods are classification (target variable is a category), regression (target and background variables are numbers), the density score (predicted value is the probability density function) (Grigorova et al. 2017). Examples of e-learning systems using prediction are AL-TESL-e-learning system (Wang and Liao 2011) for the classification of learner characteristics and learning performances and Junyi Academy (Paquette et al. 2020), which uses regression to analyze and extract information from both human annotations and usage logs.

Clustering is the process of grouping a set of objects into classes of similar objects (Asif et al. 2017). In e-learning, cluster analysis can be used to investigate similarities and differences between learners, courses, teachers, etc (Zorić 2020). Typical examples of e-learning systems using this technique are dotLRN (Köck and Paramythis 2011) for monitoring and interpreting sequential learner activities and ESURBCA (Suresh and Prakasam 2013), which uses this technique to get consistency in content delivery, quality content in learning materials, learners’ self-learning concepts, and performance improvement in their examination.

Relationship mining is the process of identifying relationships among variables in a dataset and encoding them as rules for subsequent use (Wibawa et al. 2021). The key application of this technique in e-learning is identifying relationships in learner behavior patterns and diagnosing learner difficulties (Romero and Ventura 2020). Different types of relationship mining exist, such as association rule mining (relations between variables), sequential pattern mining (temporal association between variables), correlation mining (linear correlation between variables), and causal data mining (the causal relationships between variables) (Grigorova et al. 2017). Typical examples of e-learning systems using relationship mining are eDisiplin (Man et al. 2018), which uses association rules to disclose some useful patterns for decision support, ESOG (Ougiaroglou and Paschalis 2012), which uses this technique to discover all hidden associations that satisfy some user-predefined criteria, and Blockly programming App (Shih 2018), which uses sequential pattern mining to find the sequence patterns of app functions accessed by users.

Distillation for human judgment is a process of representing data using visualization and interactive interfaces to enable humans to quickly identify or classify data features (Wibawa et al. 2021). The key application of this technique in e-learning is helping instructors visualize and analyze the ongoing activities of the learners and their use of information (Romero and Ventura 2020). A typical example of an e-learning system for distillation for human judgment is Canvas (Arnold and Pistilli 2012), which utilizes this technique to model learner and instructor usage data and to examine the relationship between these models and learner learning outcomes.

Discovery with models is a technique that uses a model developed via prediction, clustering, or by human reasoning knowledge engineering and then used as a component in further analysis (Hicham et al. 2020). It can be used to find relationships between learner behavior and learner characteristics or contextual variables, analysis of research questions in various contexts, and integration of psychometric modeling frameworks into machine learning models (Bienkowski et al. 2012). A typical example of an e-learning system using this technique is the Cisco Networking Academy (Mislevy et al. 2012), which uses it for online assessment.

Using data mining in e-learning will enable the analysis of learners’ data collected from e-learning environments to discover hidden knowledge and recognize patterns, which can then be used to improve learning.

The rapid development of computer-supported education environments and e-learning platforms provide abundant learners’ exercise data for Knowledge Tracing (KT) (Chen et al. 2018). KT is the task of modeling learner knowledge over time so that we can accurately predict how learners will perform on future interactions (Piech et al. 2015). Its working mechanism is to take observations of a learner’s performance (e.g. the correctness of the learner response in a practice opportunity) or a learner’s actions (e.g. the time he stayed for a question), and then use those to estimate the learner’s underlying hidden attributes, such as knowledge, goals, preferences, and motivational state, etc (Gong et al. 2010).

The following are the most important advantages of KT in e-learning environments:The most known techniques for KT are Bayesian knowledge tracing and deep knowledge tracing.

A family of methods called Bayesian Knowledge Tracing (BKT) uses probabilistic methods and machine learning to model learners’ skill acquisition in e-learning systems (David et al. 2016). The original BKT model was proposed by Corbett and Anderson (1994). E-learning systems that utilize BKT can predict learner performance (Qiu et al. 2011). BKT is a specific type of dynamic Bayesian network, or more precisely, of hidden Markov models consisting of observed and latent variables (performance and skills) (Schodde et al. 2017). BKT takes learner performances and uses them to estimate the student’s level of knowledge (Gong et al. 2010). BKT assumes that student knowledge is represented as a set of binary variables, one per skill (the skill is either mastered by the student or not) (Yudelson et al. 2013). They are updated based on the correctness of students’ answers to questions that test the skill under investigation; hence observations are also binary (Käser et al. 2014). Some of the typical e-learning systems based on BKT are CRYSTAL ISLAND (Rowe and Lester 2010), ASSISTment (Pardos and Heffernan 2010), edX (Pardos et al. 2013), Coursera (Wang et al. 2016), Robot language tutor (Schodde et al. 2017), and SITS (Hooshyar et al. 2018).

Deep Knowledge Tracing (DKT) proposed by Piech et al. (2015) utilizes recurrent neural network to model student learning. DKT aims to automatically trace learners’ knowledge states by mining their exercise performance data (Yang and Cheung 2018). In the DKT algorithm, at any time step, the input to recurrent neural networks is the learner performance on a single problem of the skill that the learner is currently working on (Xiong et al. 2016). As a learner progresses through an assignment, the DKT algorithm attempts to utilize information from previous timesteps, or problems, to make better inferences regarding future performance (Zhang et al. 2017b). Specifically, based on learner historical answered questions, it can predict learner performance on future questions with high accuracy (Wang et al. 2019). Typical examples of e-learning systems based on DKT are Khan academy (Piech et al. 2015) and Udacity (Kim et al. 2018).

Many other authors have developed their own models for KT through the years. Dynamic Key-Value Memory Networks (DKVMN) presented in Zhang et al. (2017a) can exploit the relationships between underlying concepts and directly output a learner’s mastery level of each concept. DKVMN is a variant of memory-augmented neural networks, a type of model that adds storage modules and corresponding read-write mechanisms based on traditional neural networks (Sun et al. 2021). Sequential Key-Value Memory Networks (SKVMN) (Abdelrahman and Wang 2019) is also a deep learning model for KT proposed by Abdelrahman & Wang, which unifies the strengths of recurrent modeling capacity and memory capacity of the existing deep learning KT models for modeling learner learning. Pavlik Jr et al. (2009) presented Performance Factors Analysis (PFA) as a new alternative to KT, where they compared these two models, and results suggested that the PFA model was somewhat superior to the KT model overall. PFA predicts learner performance based on the item difficulty and learner historical performances (Gong et al. 2010). Pu et al. (2021) presented Deep Performance Factors Analysis (DPFA) for KT, which consistently outperforms PFA and DKT and has results comparable to those of DKVMN. Chen et al. (2018) developed a new KT model named PDKT-C to estimate learners’ knowledge states better. González-Brenes et al. (2014) presented FAST (Feature Aware Student Knowledge Tracing), an efficient, novel method that allows integrating general features into KT. Pandey and Karypis (2019) proposed a self-attention-based knowledge tracing model named SAKT, which models a learner’s interaction history and predicts his performance on the next exercise by considering the relevant exercises from his past interactions. Trifa et al. (2019) proposed Mod-Knowledge, an intelligent agent that analyzes the learner interaction to trace the learner’s knowledge state using machine learning algorithms.

The main current challenge is to support KT with the psychological and behavioral aspects of the teaching process, linked specifically to learning design.

Educational data is growing rapidly as an increasing number of education systems are going online (Krikun 2017). Extensive data sets are available from learners’ interactions with educational software, and online learning (Koedinger et al. 2010; Siemens and Baker 2012). E-learning environments automatically capture system-based records of users’ activities, recording who accessed what and when (Phillips et al. 2011). The analysis of this data can improve learning models to predict the results of learners, for example, to distinguish who needs extra help or who can solve a more complex task to develop additional skills (Mamcenko and Kurilovas 2017).

Learning Analytics (LA) can be defined as the analysis of electronic learning data, which allows teachers, course designers, and administrators of learning environments to search for unobserved patterns and underlying information in learning processes (Agudo-Peregrina et al. 2014). LA has the ability to interpret the unusual behaviors of learners, recognize patterns in learning, identify potential issues or gaps, conduct appropriate interventions, and increase learners’ awareness of their actions and development (Mangaroska and Giannakos 2018; Siemens and Long 2011). It employs sophisticated analytic tools and procedures to investigate and visualize large institutional data sets in the service of improving learning and teaching (Brown 2011; Macfadyen and Dawson 2012; Shum and Ferguson 2012). The majority of techniques employed by LA are derived from EDM, but in addition to these techniques, LA also incorporates social network analysis and statistical analysis (Bienkowski et al. 2012; Romero and Ventura 2020).

EDM techniques used for LA include prediction, clustering, relationship mining, distillation for human judgment, and discovery with models. Examples of e-learning systems that use some of the EDM techniques are Desire2Learn, Shehata and Arnold (2015) which uses prediction to improve learner success, retention, completion, and graduation rates, and Blackboard Vista (Arnold and Pistilli 2012), which employs distillation for human judgment for LA.

Social network analysis records learners’ online interactions to gauge their participation and engagement level (Czerkawski 2015). The purpose of social network analysis is to determine and understand the relationships between learners in a network environment such as discussion forums or social media (Wibawa et al. 2021). A typical example of an e-learning system that uses this technique is OSBLE+ (Olivares et al. 2019), which uses social network analysis to determine whether exposure to automated interventions would positively affect the relationship among learners over time.

Statistical analysis is used for the analysis and interpretation of quantitative data for decision-making (Leitner et al. 2017). By using statistical analysis in e-learning environments, we can count the number of visits, analyze mouse clicks, and calculate time spent on tasks (Khalil and Ebner 2015). A typical example of e-learning system using this technique is OU Analyse (Herodotou et al. 2020), which uses a variety of advanced statistical approaches to identify learners who are at risk of failing their studies.

Visualization of data obtained using these techniques is mainly done using learning analytics dashboards. For learners and teachers alike, they should be able to see a visual representation of their activities and how they relate to those of their classmates or other participants in the learning process (Duval 2011).

A Learning Analytics Dashboard (LAD) is a single display that aggregates different indicators about learners, learning processes, and learning contexts into one or multiple visualizations (Schwendimann et al. 2016). LAD provides educators and learners with a comprehensive snapshot of the learning domain (Ramaswami et al. 2022). LAD display insights, trends, and changes in data over time, mainly using different types of charts and various widgets. They present data clearly and efficiently, and they are very intuitive for both teachers and learners. The main goal of LAD for educators is to assist them in getting learners’ feedback either during or after lectures so they can use this information to modify and adjust their instructions, lessons, guidance, and tutoring (Verbert et al. 2014). Also, teachers can use this feedback for checking each learner’s learning progress and providing proper interventions (Kim et al. 2016).

The LAD and OLM have similar goals. Both the LAD and OLM aim to make data available to help learners interpret aspects of their learning (Kay and Bull 2015). OLM shows the learner model to users, with some visualization, to assist their self-regulated learning by, for example, helping prompt reflection, facilitating planning, and supporting navigation (Guerra-Hollstein et al. 2017). Even though OLM and LA are similar areas concerning the visualization of learners’ data (in general), they are different in that OLM is more inclined toward visualization of the learner’s level of knowledge (progress), difficulties, and misconceptions, while LA is rather more based on results towards prediction, recommendation, and also semantic aggregation (Hooshyar et al. 2019; Kay and Bull 2015). In addition, a significant difference is that OLMs are grounded on work in “student modeling”, “learner modeling”, and even the broader “user modeling”, whereas dashboards are more broadly grounded in data-driven decision-making, which often includes goals, stakeholders, and decision-making outside of the context of the learner model (Bodily et al. 2018).

LA has the potential for a tangible positive impact on learner learning by supporting learning strategies and tactics (Knight et al. 2020). Learning strategies can be defined as the behaviors of a learner that are intended to influence how the learner processes information (Mayer 1988; Gasevic et al. 2017). A learning tactic is a single or a very short sequence of operations a learner applies to information (Winne and Marzouk 2019). Traditionally, learning strategies and tactics are usually discovered using surveys, questionnaires, or think-aloud protocols. In e-learning, LA has the potential to detect and explain characteristics of learning strategies and tactics through analysis of trace data and communicate the findings via feedback (Matcha et al. 2019b). Such LA approaches provide a direct analysis of the users’ “actual” behavior in lieu of the learners’ perception, and recall of events (Jovanović et al. 2017).

After collecting the learners’ data and discovering their learning strategies and tactics via LA, learning designers use this valuable information for learning design. Learning design can be defined as an application of a pedagogical model for a specific learning objective, target group, and a specific context or knowledge domain (Koper and Olivier 2004). Learning designers use summative, real-time, and predictive LA to increase the quality of the curriculum, materials, scaffolds, and assessments (Ifenthaler 2017). Learning design and LA can help teachers in designing quality learning experiences for their learners, evaluate how learners are learning within that intended learning context, and support personalized learning experiences for learners (Lockyer and Dawson 2011).

LA can be a powerful tool in e-learning environments. Using LA in e-learning can assist in identifying potential issues or gaps in learners, and then instructors can use this information to optimize learning and improve learners’ success.

In traditional education, most assessments present the same set of questions to all learners. While such examinations are relatively simple to develop, they do not reveal what each learner is genuinely capable of. There may be learners who know more than their test results display. Alternatively, learners may answer some questions poorly but might demonstrate a higher level of knowledge if given more straightforward questions. If all learners are asked the same questions, the information provided by the results may be limited.

In contrast, adaptive assessment is adjusted to learner responses by analyzing their range of skills and proficiency levels and defining a tailored learning route for a specific learner. Such adaptive assessment systems are also known as computerized adaptive testing, which represents a special case of computer-based testing where each examinee takes a unique test that is tailored to their ability level (Triantafillou et al. 2008). With the use of adaptive testing, the evaluation system is more accurate in measuring the ability of the learner. It can accommodate the diversity of learner capabilities to provide learning materials for the system according to the level of learner’s proficiency (Kustiyahningsih and Cahyani 2013). A computer-based adaptive assessment provides highly accurate results that can recognize the mastered competencies of each learner, determine educational requirements, track educational advancement throughout time, and place learners in educational programs that suit them (Raman and Nedungadi 2010).

The most valuable benefits of adaptive assessments in online learning environments are:The most common intelligent techniques used in e-learning software development to match learners’ proficiency level are item response theory, Elo rating algorithm, and TrueSkill.

Item Response Theory (IRT) is a conceptual framework that is based on basic measurement concepts using statistical and mathematical tools (Salcedo et al. 2005). IRT has roots in psychometrics and is concerned with accurate test scoring and the development of test items. It is usually used in the adaptive assessment systems domain to calibrate and evaluate items in a test and to score learners’ abilities, attitudes, or other latent traits (Dahl and Fykse 2018). IRT models enable estimation of the skill level of each learner and its evolution over time, as well as the difficulty and the discrimination of each question (Benedetto et al. 2020). Some typical examples of e-learning systems that implement adaptive assessment based on IRT are UZWEBMAT (Özyurt et al. 2014), MISTRAL (Oppl et al. 2017), Amrita Learning (Gutjahr et al. 2017), SIETTE (Conejo et al. 2018), UTS (Bernardi et al. 2018), English vocabulary learning system (Chen et al. 2019), SamurAI (Uto et al. 2019), PEL-IRT (Ferjaoui and Cheniti Belcadhi 2020), and Lexue 100 (Jia and Le 2020).

Elo rating algorithm represent an alternative to IRT. The Elo rating algorithm, developed to measure player strength in chess tournaments, has also been applied for educational research and used to measure learner ability (Pankiewicz and Bator 2019). To adapt the concept of a chess game to the educational measurement, a learner is considered a player, an item is considered an opponent, and a correctly answered item is considered a win for the learner (Park et al. 2019). Some typical examples of e-learning systems that implement adaptive assessment based on Elo rating algorithm are ProTuS (Mangaroska et al. 2018), Math Garden (Brinkhuis et al. 2018), Matistikk (Dahl and Fykse 2018), and ACT Academy (Yudelson et al. 2019).

TrueSkill is another approach for adaptive assessment, which can be viewed as a generalization of the Elo rating system used in chess (Herbrich et al. 2006). TrueSkill was developed for ranking players in video games. By interpreting problem-solving as a match between learner and problem, it is possible to use the TrueSkill rating system to estimate the ability of learners to solve a series of issues in the e-learning environment (Lee 2019). A typical example of an e-learning system that implements adaptive assessment based on TrueSkill is APACTS (Kawatsu et al. 2017).

Adaptive assessment can change the difficulty level based on a learner’s responses, making it more efficient, targeted, and precise than traditional tests. They are, however, extremely complex, time-consuming, and resource-intensive to create.

Personalized learning is a learning experience in which the pace of learning and the instructional approach are tailored to each learner’s specific needs (Mikić et al. 2022). The pace of learning, sequence, technology, instructional strategy, instructional content, and other aspects of personalized learning may all vary based on learner needs. This more tailored education aims to provide relevant learning activities motivated by their interests and are frequently self-initiated.

Personalized e-learning systems customize various features of online learning, such as the user interface, learning content, or activities. These features can be personalized based on different factors such as learners’ prior knowledge, preferences, habits, behavior, etc. AI shows its true potential when designing personalized training, providing a wide range of methods and techniques that could be applied to customize the learning process in online environments. The most commonly used personalization methods in e-learning today are:With the implementation of personalized learning, AI will assist learners in achieving academic success while also helping teachers in becoming more effective.

Intelligent agents as a conception has been around since the early years of e-learning. These so-called assistant programs are located inside the e-learning system, and they make the whole learning process more active to suit the learner’s demands. Intelligent agents are used to collect, process, and analyze learners’ data to understand and optimize the learning environment (Kotova 2017). In general, intelligent agents can monitor behavior, evaluate learners’ performance and the importance of the way of transmitting recommendations, and improve the quality of learning (Saeidi Pour et al. 2017). Agents can perform repetitive tasks, remember things learners forgot, intelligently summarize complex data, learn from learners and even make recommendations to them (Meleško and Kurilovas 2018a).

The utilization of agents in e-learning systems assists in referring to many of the restrictions of these systems by promoting the design and distribution of e-learning objects that match learner preferences. These agents observe the learner’s responses to the introduced learning object to provide convenient content for the learner’s abilities. Agents are concentrated task-resolving units with defined task margins and can perform independently without directions.

Up-to-date extraction of learning contents, easy, comfortable, fast contact between teacher and learner, and supervision over the teacher’s and learner’s performance during the learning process are considered advantages provided by a system based on intelligent agents in e-learning (Fasihfar and Rokhsati 2017).

Some typical examples of autonomous intelligent agent systems are eTeacher (Chung 2015), which tracks the learning activities and performance of learners and provides them with personalized content, and Mod-Knowledge (Trifa et al. 2019), which is responsible for analyzing the learner’s interactions to trace their knowledge state.

A multi-agent system is a collection of autonomous agents that work together to solve problems that are beyond the capabilities of individual agents (Xu et al. 2014). They can be described as an assembly of agents with their trouble-fixing capabilities.

Examples of multi-agent systems are MAS-PLANG (Asselman et al. 2018), which provides content, navigation strategy, and navigation tools according to the learners’ learning habits, EMASPEL (Bokhari and Ahmad 2014), which takes the facial expression of the learner and provides help accordingly, ALLEGRO (Asselman et al. 2017), which manages the content for learners, F-SMILE (Alkhatlan and Kalita 2019), used to generate default assumptions about learners, ISABEL (Aguilar et al. 2015), where the idea is to divide the learners in groups with similar profiles and where each group is managed by a tutor agent, IAELS (Tsai et al. 2012), for improving learners’ learning capabilities, MIPITS (Lavendelis 2015), used to provide different types of tasks and adapt tasks to the needs of individual learners, ADOPT (Ajroud et al. 2021), whose agents analyze the traces left by the learner, calculate various indicators, and propose the most suitable adaptations for the learner, PowerChalk (Rosado et al. 2015), whose agents are used for identification in order to get individual user profiles and to create educational content, and MetaTutor (Trevors et al. 2014), which interacts with the learner’s prior domain knowledge to affect their note-taking activities and subsequent learning outcomes.

Personalized approaches based on intelligent agent technology imply that each learner has its own personal (pedagogical) agent, tutor respectively, that directs learner towards learning (Kuk et al. 2016). These agents are in the form of a virtual character or simple chat interface equipped with AI that can support the learners’ learning process and use various instructional strategies in an e-learning environment (Martha and Santoso 2019).

In e-learning, there are two types of these agents: pedagogical agent and conversational agent. The difference between these two is that a pedagogical agent will merely hold a monologue, give hints and solutions, while a conversational agent can engage in a dialogue with the learner (Davis 2018; Wellnhammer et al. 2020). Conversational agents use natural language processing, which enables them to understand text and spoken words and converse with humans in a natural way (Van Pinxteren et al. 2020). The main goals of these agents are to motivate and guide learners through the learning process by asking questions and proposing solutions (Klašnja-Milićević et al. 2017).

Typical examples of e-learning systems with pedagogical agents are ISLA (Mondragon et al. 2016) with pedagogical agent Jessi, who is capable of detecting the affective state of an autistic child in mathematical learning, ATS (Thompson and McGill 2017), in which pedagogical agent Jean responds to the affective state of the learner, CALL system (Carlotto and Jaques 2016), with pedagogical agent Patti who explains the material on the content screens and reacts to learners’ exercise results, APLUS (Matsuda et al. 2014) has a pedagogical agent named SimStuden, who helps learners learn to solve algebraic equations by tutoring, and a second agent, Mr. Williams, on whom learners can click to ask for help, and the EC Lab (Osman and Lee 2014), which also has two pedagogical agents, Professor T., who gives accurate information and explains new concepts to the learners, and Lisa, who learns together with the learners and provides motivation and encouragement to the learners to complete the tasks and exercises in the module.

Typical examples of e-learning systems with conversational agents are Geranium (Griol et al. 2014), with conversational agent Gera, which poses questions to the learners that they must answer either orally or using the graphical interface, and Curiosity Notebook (Chhibber and Law 2019), which supports conversational interaction between learners and AI agents.

Including agents in e-learning will push the learning process to a more functional level. With intelligent agents, e-learning systems have the chance to learn and use their acquired knowledge to fulfill their goals (Fasihfar and Rokhsati 2017).

Previous chapters represent the overview of different e-learning processes and intelligent techniques that are used to support them. However, most popular intelligent techniques found their way into e-learning by offering multiple different roles, most notably artificial neural networks, Bayesian network, fuzzy logic, decision tree, and hidden Markov model.

Artificial neural networks are a powerful class of machine learning algorithms that learn complex patterns from data using collections of simple trainable mathematical functions (Hassan and Hamada 2017). In e-learning systems, they are mostly used for predicting a learner’s performance (Şuşnea 2010), analyzing learners’ online interactive behaviors (Poitras et al. 2019), recognizing the cognitive state of learners (Bhattacharya et al. 2018), detecting human expression using psychological signals (Dogmus et al. 2014), or predicting difficulties a learner will experience in a digital design course (Hussain et al. 2019).

A Bayesian network represents some conditional dependency of the random variables by the directed acyclic graph and conditional probability tables (Kondo and Hatanaka 2019). It is a combination of AI, probability theory, graph theory, and decision theory (Wang et al. 2020). Bayesian networks can learn and represent uncertain and coarse knowledge (Zhang et al. 2021a). It has significant advantages in dealing with uncertainty, mining the correlations among observable quantities, latent variables, and unknown parameters (Jiang et al. 2018). In e-learning systems, it is mostly used for detecting the learning habits of a learner (Leka and Kika 2018), managing learner models (Nguyen and Pham 2011), evaluating the teaching performance of university teaching assistants in an e-learning session (Xu et al. 2016), adaptively supporting students in learning environments (Eryılmaz and Adabashi 2020), or for the learners’ acquisition of problem-solving skills in computer programming (Hooshyar et al. 2018).

Fuzzy logic is an extension of the traditional set theory, as statements can be partial truths, lying in between absolute truth and absolute falsity (Bajaj and Sharma 2018). In e-learning systems, it is mostly used for evaluation and assessment of learners’ tasks (Jurado et al. 2014), checking the quality of the proposed solutions by the learners and monitoring the collaboration process (Ortega 2021), identification of the alterations on the state of learners’ knowledge level and making dynamic decisions on how the teaching syllabus is presented to the learner to fit his/her personal needs (Chrysafiadi and Virvou 2014), identifying learners’ preferred learning strategies and knowledge delivery needs that revolve around characteristics of learners and the existing knowledge level in generating an adaptive learning environment (Almohammadi et al. 2016), or for describing, defining, and modifying the uncertainty in a student’s behavior (Katz et al. 1994).

A decision tree is a tree in which each branch node represents a choice between several alternatives, and each leaf node represents a decision (Bajaj and Sharma 2018). It is frequently used to acquire information and then use it to predict outcomes, and therefore facilitates decision-making (Al Karim et al. 2021). In e-learning systems, it is mostly used for analyzing data and sharing the results (Akyuz 2020), building a learner model that predicts each learner’s final learning state (Uto et al. 2019), providing personalized learning paths for optimizing the performance of creativity (Lin et al. 2013), extracting and highlighting important information from learners’ data (Karkar et al. 2020), or for the classification of learners’ learning mistakes (Faeskorn-Woyke et al. 2020).

A hidden Markov model is a collection of unobserved states which follow the rules of the Markov property in a situation where the interrelationship between the true observation and an unobserved state takes birth from a probability distribution (Almohammadi et al. 2017). In e-learning systems, it is mostly used for the adaptability of course content according to the learner’s performance in the pre-test and post-test (Rani et al. 2017), predicting students’ navigation actions (Homsi et al. 2008), measuring motivation and prior knowledge of the learners (Van Seters et al. 2012), enhancing navigational abilities and easing the search for suitable learning content for the visually impaired (Azeta et al. 2014), or automatic recognition of e-learning activities based on the mouse movement of learners (Elbahi et al. 2016).

AI and other emerging technologies are transforming modern e-learning. AI integrated into e-learning solutions aids in creating custom-tailored learning paths, personalizing online courses, providing relevant materials to appropriate learners, analyzing the content to improve learner engagement, and automating the learning and grading processes. Without a doubt, AI has limitless potential in the e-learning industry, where technology integration can reshape the way we learn.