F-SMILE: An Intelligent Multi-Agent Learning Environment Maria Virvou, Katerina Kabassi Department of Informatics, University of Piraeus,
[email protected],
[email protected] Abstract This paper describes a multi-agent, intelligent learning environment. The system is called F-SMILE and is meant to help novice users learn how to manipulate the file store of their computer. F-SMILE consists of a Learner Modelling (LM) Agent, an Advising Agent, a Tutoring Agent and a Speech-driven Agent. The LM Agent constantly observes the learner and in case it suspects that the user is involved in a problematic situation it tries to find out what the cause of the problem has been. This is done by searching for similar actions to the one issued by the learner which would have been more appropriate with respect to the hypothesised learner’s goals. These alternative actions are sent to the Advising Agent, which is responsible for selecting the most appropriate action to be suggested to the particular learner. The selection of the best alternative action is based on the information about the learner that the LM Agent has collected and a cognitive theory. In case the problem of the user was due to lack of knowledge, the Tutoring Agent is activated in order to generate adaptively a lesson appropriate for the particular user. When the advice and the corresponding lesson are ready, they are sent to the Speech-driven Agent, which is responsible for rendering the interaction with the user more human-like and user-friendly.
1. Introduction Software complexity often discourages novice users to learn how to use effectively a software system. Therefore, learning environments should monitor the users’ progress, while they are actively engaged in problem-solving activities, and provide them with feedback in a manner that contributes to achieving the twin goals of learning effectiveness and learning efficiency [8]. In response to these requirements, software agents play an important role in the human-computer interaction and in the coordination of the internal processes of the system [1]. F-SMILE (File-Store Manipulation Intelligent Learning Environment) is an intelligent learning environment for novice users of a GUI (Graphical User Interface) that manipulates files, such as the Windows 98/NT Explorer [10]. The system offers a protected
environment to novice users who may work as they would normally do while the system reasons silently about their actions. In case the system diagnoses a problematic situation it offers spontaneous advice. In addition if the problem has been due to the user’s lack of knowledge F-SMILE provides adaptive tutoring. However, the development of a system that intervenes appropriately requires a mechanism that infers the user’s intentions and diagnoses possible problems. Diagnosis should be conducted in a non-invasive way because interrupting users to determine their current intentions and to disambiguate their misconceptions may become annoying. Therefore, F-SMILE has assigned agents to observe users while they are actively engaged in their usual activities and provide spontaneous advice in case this is considered necessary. Agents have been widely used in learning environments in order to play different roles (e.g. [11], [17]) or perform certain tasks, such as capturing the user’s characteristics (e.g. [2], [14]). However, the majority of agent based architectures consist of a single agent [18]. The main disadvantage of such an approach is that the agent’s knowledge, computing resources and perspective is limited. These problems are addressed by multi-agent systems. The design of individual agents within a multi-agent system has the advantage of being independent of the design of other agents as long as each agent conforms with an agreed upon protocol and ontology. This contributes to the breakdown of complexity significantly [4]. It is widely agreed that basing decisions on an accurate cognitive model of the user is important for effective prediction of user intent [13]. To this end, we have adapted and implemented a part of a theory called Human Plausible Reasoning theory [3]. Human Plausible Reasoning theory (HPR) is a domain-independent theory originally based on a corpus of people’s answers to everyday questions. F-SMILE uses HPR in error diagnosis and for the generation of more human-like advice.
2. Multi-agent Architecture of F-SMILE F-SMILE (File-Store Manipulation Intelligent Learning Environment) is an intelligent learning
environment that is addressed to novice learners who learn how to use file-store manipulation programs such as Windows 98 Explorer [10]. The system constantly reasons about every learner’s action and provides spontaneous advice in case this is considered necessary. Advice is provided to learners who have made an error with respect to their hypothesised intentions. F-SMILE is based on a multi-agent architecture. Multi-agent systems are capable of solving problems that are too complicated for a single agent to solve because of its resource limitations [18]. F-SMILE’s architecture consists of four agents, namely, Learner Modeling (LM) Agent, Advising Agent, Tutoring Agent and Speechdriven agent, and the domain representation component. The architecture of F-SMILE can be seen in Figure 1.
alternatives is based on inference patterns provided by the Human Plausible Reasoning theory. As soon as the alternative actions are generated, they are sent to the Advising Agent, which is responsible for selecting the alternative action that the learner was more likely to have intended. For this purpose, the Advising Agent also uses long-term information about the learner’s habits, level of knowledge and proneness to errors, which is stored in the long-term learner model. Furthermore, in case the LM Agent thinks that the learner's misconception was due to the learner's lack of knowledge, it informs the Tutoring Agent accordingly. The Tutoring Agent is responsible for forming an adaptive presentation of the lesson to be taught to the learner. Both the Advising Agent and the Tutoring agent send their results to the Speech-driven Agent. The Speechdriven Agent is responsible for presenting the information in a unified and easy to access fashion. In order to make the interaction more natural and enjoyable, an animated, speech-driven Agent is used to present the system’s advice to the learner. As Rist et al. [16] point out, such characteristics add expressive power to a system’s presentation skills.
3. The Learner Modeling Agent
Figure 1: F-SMILE's architecture Every time the learner issues a command, the LM Agent reasons about it with respect to its expectations about the learner’s goals. If an action contradicts its expectations, the LM Agent is responsible for diagnosing the cause of the learner’s error. This is done by searching for similar alternative actions to the one issued that the learner may have intended to issue instead of the one issued which was problematic. The search for similar
The Learner Modelling (LM) Agent captures the cognitive state, as well as the characteristics of the learner and identifies possible misconceptions. This approach is similar to [9]. Generally, the LM Agent observes the learners while they are actively engaged in their usual activities, maintains and manages the learner profiles and provides relevant information whenever other agents request it. In case the LM Agent suspects that the learner is involved in a problematic situation, it performs error diagnosis. For this purpose, it employs an analysis engine to derive new ‘facts’ about the learner and to respond to queries from other agents. The analysis engine is based on a limited goal recognition mechanism and HPR. In particular, the generation of hypotheses is based on HPR, which is a theory about human plausible reasoning. Prior to F-SMILE, HPR had also been successfully used for simulating the users’ reasoning in a help system for a different domain [19]. However, in F-SMILE it has additionally been used to provide a simulation of human tutors’ reasoning when they form the advice to be given to students. HPR detects the generalisation, specialisation, similarity, dissimilarity relationship between a question and the knowledge retrieved from memory and drives the line (type) of inference. However, this procedure usually results in the generation of many alternative actions. Therefore, all alternative actions are sent to the Advising Agent, who is responsible for selecting the one that the learner was more likely to have intended. The LM Agent also uses an adaptation of the certainty parameters of
HPR to capture long-term information about the learner. The certainty parameters that are used by the LM Agent are the degree of typicality, the degree of frequency and the degree of dominance. Certainty parameters will be discussed in more detail in the next section. Although agents are successful in being able to learn their learner’s behaviour and assist them, a major drawback of these systems is the fact that they require a sufficient amount of time before they can be of any use [6]. A solution to this problem may be the incorporation of stereotypes in the system [15]. The LM Agent in FSMILE uses a novel combination of HPR with a stereotype-based mechanism in order to generate default assumptions about learners until it acquires sufficient information about each individual learner. Learners are classified into three major classes according to their level of knowledge and to two classes according to their degree of carelessness. In particular, the classes concerning the level of knowledge are ‘absolute beginners’, ‘intermediate’ and ‘advanced’. The classes concerning the learners’ degree of carelessness are ‘careless’ and ‘careful’. In addition to stereotypes, the LM Agent is also constantly collecting information about a particular learner’s behaviour and errors and updates the learner’s individual learner model. The percentage of information acquired by the stereotype diminishes as the percentage of acquisition from the individual learner model increases. In case a conflict appears, the LM Agent always favours information provided by the individual learner model. The LM Agent collects, combines and evaluates information in order to identify the learner's errors. For this purpose, the LM Agent generates alternative actions that the learner may have meant to issue instead of the ones issued, which were problematic. If the LM Agent does not manage to identify any error then the learner is not interrupted at all and his/her actions are executed normally. In cases where the LM Agent has identified lack of knowledge of the learner on something then not only does it generate alternative commands but it informs the Tutoring Agent that the learner needs additional tutoring on the particular subject, as well.
4. The Advising Agent The Advising Agent tries to simulate human tutors' reasoning by using an adaptation of the certainty parameters introduced in HPR. The main task of the Advising Agent is to select the most appropriate advice for the learner based on information about the learner and possible errors that the LM Agent has sent. The adaptation of the certainty parameters used by the Advising Agent are presented below:
• • • • •
The degree of similarity (σ) is used to calculate the resemblance of two commands or two objects. The typicality (τ) of a command represents the estimated frequency of execution of the command by the particular learner. The degree of frequency (ϕ) of an error represents how often a specific error is made by a learner. A learner’s most common errors can be recognised by the dominance (δ) of an error in the set of all errors of the learner. Finally, all the parameters presented above are combined to calculate a degree of certainty related to every alternative command generated by the LM Agent. This is the degree of certainty (γ) and represents the degree of the agent's certainty that the learner intended the alternative command generated.
In order to calculate the degree of certainty for each alternative action, the Advising Agent multiplies each parameter by a weight. The weight is determined with respect to how important the particular certainty parameter is with respect to human tutors’ reasoning. An evaluation of the design of the advice generator of the system [21] revealed that the most important criterion of a human expert when evaluating an alternative action, was the similarity of that action to the one issued by the learner, because learners usually tend to tangle up actions or objects that are very similar. The second most important criterion that human experts used was whether a particular learner’s error was the most frequent error of all errors that this learner made. Moreover, the frequency the learner made such an error while interacting with the system, was to be taken into account even if this error was not the most frequent one for this particular learner. The last thing that human tutors took into account was whether a learner used that particular action quite often or not. In view of the above the formula for the calculation of the degree of certainty was formed in such a way that the reasoning of F-SMILE was close to human experts’ reasoning: γ = 0.4 * σ + 0.3 * δ + 0.2 * ϕ + 0.1 * τ
(1)
5. The Tutoring Agent The Tutoring Agent uses the information about a particular learner and his/her misconception identified by the LM Agent, to adapt the content of the lesson, links and examples being presented to that learner. It is among the goals of F-SMILE to provide the ‘right’ pieces of information in the ‘right’ way and at the ‘right’ time. For this purpose, the Tutoring Agent uses adaptive hypermedia techniques to protect learners from
information overflow, to help them find easily a required piece of information and to help them understand new pieces of knowledge that are being taught. In particular, these techniques use information about a particular learner, represented in the learner model, to adapt the lessons presented to that learner. In F-SMILE, adaptive presentation techniques to the learner are used to present examples of use of an unknown command in the context of the learner’s own file-store. Therefore, the Tutoring Agent generates examples dynamically so that it may use the names of the particular learner’s existing files and folders. Moreover, the Tutoring Agent uses adaptive link annotation techniques to present to the learner other parts of knowledge that are believed to be of interest to the learner for the particular case.
6. The Animated, Speech-driven Agent The user interface of F-SMILE uses an animated, speech-driven character. Such characters provide an entertaining and emotional function which may help to lower the ‘getting started barrier’ for novice learners of computer applications. In addition, such characters improve the effectiveness of the system by engaging and motivating learners [5] and prospective learners in an educational application [12]. Furthermore, Walker et al. [22] using an analysis of subjects’ responses to a synthesised talking head showed that subjects reacted better to spoken than written information. Therefore, an animated, speech-driven agent is employed to emulate more familiar communication styles and make the interaction with the learners more natural. The Speech-driven Agent is responsible for the overall communication with the learner. This usually involves the collection of the learner’s queries and the presentation of advice in case the learner is diagnosed to have been in a problematic situation. However, the particular agent does not contain any further reasoning mechanisms.
7. Example operation of F-SMILE In this section we present a simple example of the system's operation. The learner's initial file store state is shown in figure 2. The learner, called Helen, had a floppy disk that she wanted to format. 1. 2. 3. 4.
create_new_folder_in(C:\My Documents\) rename(C:\ My Documents\NewFolder\, C:\ My Documents\floppy\) cut(A:\essayGR.txt) copy(C:\ My Documents\floppy\) LM Agent's reasoning: Suspect Action. Suggestion: paste(C:\My Documents\floppy\)
5. 6.
paste(C:\My Documents\floppy\) format(A:\)
Figure 2: The learner’s initial file store state The learner created a new folder in the hard disk in the folder C:\My Documents\ which she called ‘floppy’ (action 1-2). Her intention was to move in the new folder some of the contents of the floppy disk. Then she tried to move the file ‘essayGR.txt’ from the floppy disk to the newly created folder by using a ‘cut’ command (action 3). However, in the fourth action she used a ‘copy’ command. The LM Agent found this action suspect because if it was executed it would delete the content of the clipboard before this was used anywhere. Therefore, the LM Agent tried to generate alternative actions that the learner may have meant to issue instead. The Advising Agent evaluated the possible alternatives and found that the learner probably meant to issue: paste(C:\My Documents\floppy\) because ‘copy’ and ‘paste’ are both actions that involve the clipboard, therefore they are considered similar. Moreover, the action paste(C:\My Documents\floppy\) is considered better than the one issued because it uses the content of the clipboard and it gives contents to the newly created folder ‘C:\My Documents\floppy\’ which is empty. Therefore, the advice was formed and the Tutoring Agent was informed that the learner needed additional tutoring in the domain of copying and moving objects. The content of the response of these two agents is sent to the Speech-driven Agent, who informs the learner that she probably wanted to issue the action 'paste(C:\My Documents\floppy\)' and that she should take additional tutoring on the domain of copying and moving objects. However, the learner was not obligated to follow this advice or take the lesson, which was made for her. She could execute her initial action or issue a completely new one. In the particular example, the learner found the system’s advice very helpful and, therefore, adopted its suggestion in action 5. Then, in action 6, the learner formatted the floppy disk, which was her final goal. In case the learner had used a standard file manipulation program, her error in command 4 would not have been recognised and the learner would have formatted the
floppy disk and would have lost the useful file ‘essayGR.txt’.
8. Conclusions In this paper, we have described a multi-agent learning environment. The system helps learners learn how to operate their file store. Learners are monitored while working in a protected mode. Meanwhile, the system tries to identify problematic situations and diagnose the cause of the problem, so that it can offer appropriate advice. Novice learners can benefit from the system’s advice and adaptive tutoring facilities so that they may learn from their own errors. The reasoning process is based on a simulator of human error generation, which is based on a domain independent cognitive theory. The main focus of this paper has been on the architecture of the system and the information exchanged between the agents. The design of the system presented builds on the experiences of an earlier project [20], [21] that did not have an agent-based architecture. Therefore, the advantages of using such an approach were identified. The multi-agent architecture is open and extensible. Furthermore, using the multi-agent approach has the advantage of the decomposition of the intelligence of the system into units with autonomy (agents) that simplifies the task of designing, building and refining the individual agents.
9. References [1] L. Aroyo and P. Kommers, “Preface-Intelligent Agents for Educational Computer-Aided Systems”, Journal of Interactive Learning Research, Special Issue on Intelligent Agents for Educational Computer-Aided Systems, 1999 Vol. 10, No 3/4, pp. 235-242. [2] S. Belkada, A.I. Cristea & T. Okamoto, “Measuring Knowledge Transfer Skills by Using Constrained-Student Modeler Autonomous Agent”, Proceedings of the IEEE International Conference on Advanced Learning Technologies (ICALT 2001), IEEE Computer Society, 2001, pp. 375-378. [3] A. Collins and R. Michalski, “The Logic of Plausible Reasoning: A core Theory”, Cognitive Science, Elsevier Science, The Netherlands, 1989, Vol. 13, pp. 1-49 [4] S. El-Beltagy, D. De Roure and W. Hall, “A Multiagent system for Navigation Assistance and Information Finding. In Proceedings of the 4 International Conference on the Practical Applications of Intelligent Agents and Multi-Agent Technology, 1999, pp. 281-295. [5] W.L. Johnson, J.W. Rickel, and J.C. Lester, “Animated Pedagogical Agents: Face-to-Face Interaction in Interactive Learning Environments”, International Journal of Artificial Intelligence in Education, 2000, Vol. 11, pp. 47-78.
[6] Y. Lashkari, M. Metral and P. Maes, “Collaborative Interface Agents”, Proceedings of the 12th National Conference on Artificial Intelligence, 1994, pp.444-449. [7] V. Lesser (1995) Multiagent Systems: An Emerging Subdiscipline of AI. ACM Computing Surveys, 1995, Vol. 27, No 3, pp. 340-342. [8] J.C. Lester, B.A. Stone and G.D. Stelling, “Lifelike Pedagogical Agents for Mixed-Initiative Problem Solving in Constructivist Learning Environments”, User Modelling and User Adapted Interaction, 1999, Vol. 9, pp. 1-44. [9] G. McCalla, J. Vassileva, J. Greer and S. Bull, “Active Learner Modeling”, Fifth International Conference on Intelligent Tutoring Systems, 2000. [10] Microsoft Corporation, Microsoft® Windows® 98 Resource Kit, Microsoft Press, 1998. [11] T. Mengelle & C. Frasson, “A Multi-Agent Architecture for an ITS with Multiple Strategies”, CALISCE 96: International Conference on Computer Aided Learning & Instruction in Science Engineering, 1996, pp. 96-104. [12] M. Moundridou & M. Virvou, “Evaluating the persona effect of an interface agent in an intelligent tutoring system”, Journal of Computer Assisted Learning, 2002, Vol. 18, No. 2, [13] R. Opperman, “Adaptively supported adaptivity”, International Journal of Human-Computer Studies, 1994, Vol. 40, pp. 455-472. [14] C. O'Riordan and J. Griffith, “A Multi-Agent System for Intelligent Online Education”, Journal of Interactive Learning Research, Special Issue on Intelligent Agents for Educational Computer-Aided Systems, 1999, Vol. 10, No 3/4, pp. 263-274. [15] E. Rich (1999) Users are individuals: individualizing user models, International Journal of Human-Computer Studies, 51, 323-338. [16] T. Rist, E. Andre and J. Muller, “Adding Animated Presentation Agents to the Interface”, International Conference on Intelligent User Interfaces, 1997, pp. 79-86. [17] K. Solomos and N. Avouris, “A Multi-Agent System of Intelligent Online Education”, Journal of Interactive Learning Research, Special Issue on Intelligent Agents for Educational Computer-Aided Systems, 1999 Vol. 10. No 3/4, pp. 243-262. [18] K. Sycara, “Multi-Agent Systems”, AI Magazine, 1998, Vol. 19, No. 2. [19] M. Virvou and B. Du Boulay, “Human Plausible Reasoning for Intelligent Help”, User Modeling and UserAdapted Interaction, Kluwer Academic Publishers, The Netherlands, 1999, Vol. 9, pp. 321-375. [20] M. Virvou and K. Kabassi, “An Intelligent Learning Environment for Novice Users of a GUI. In Lecture Notes in Computer Science (Intelligent Tutoring Systems, Proceedings of the 5th International Conference on Intelligent Tutoring Systems), Springer, Berlin, 2000. Vol. 1839, pp. 635-644. [21] M. Virvou and K. Kabassi “Evaluation of the advice generator of an intelligent learning environment”, Proceedings of the IEEE International Conference on Advanced Learning Technologies (ICALT 2001), IEEE Computer Society, 2001, pp. 339-342. [22] J.H. Walker, L. Sproull and R. Subramani, “Using a Human Face in an Interface: Proceedings of ACM Conference on Human Factors in Computing Systems CHI’94, 1994, pp. 8591.
