Potentials of Learning Analytics - Amit Bahl

The term Learning Analytics has emerged to describe the process in understanding the behaviours of learning process from the data gathered from the interactions between the learners and contents. The term can be defined as as the measurement, collection, analysis and reporting of information about learners and their contexts for the purposes of understanding and optimizing learning. Another simple definition states “learning analytics is about collecting traces that learners leave behind and using those traces to improve learning”. A number of authors have considered the importance and impact of learning analytics in the future of education. In their view, the field of learning analytics is the confluence of knowledge drawn from related disciplines such as educational psychology, learning sciences, machine learning, data mining and human-computer interaction (HCI). 

Many studies have been reported the positive contributions of learning analytics. The encouraging results confirm that if properly used, learning analytics can help instructors to identify the learning gaps, implement intervention strategies, increase students’ engagement and improve the learning outcomes. From the abstract and citation database of peer-reviewed literature, its identified that case studies report empirical findings on the application of learning analytics in higher education. A total of 43 studies were selected for in-depth analysis to discover the objectives, approaches and major outcomes from the studies. The study classifies six aspects that learning analytics can support to improve the education process. These are (i) improving student retention, (ii) supporting informed decision making, (iii) increasing cost-effectiveness, (iv) understanding students’ learning behavior, (v) arranging personalized assistance to students, and (vi) providing timely feedback and intervention. These aspects are not to consider in separate entity, but are inextricably linked.

(i) Improving student retention 

In educational settings detecting early warning signs for students who are coping with their study can be an advantage for the instructors. The issues and problems that students are facing may varies from social and emotional issues to academic matters or other factors that may lead to giving-up from the study. Those students can be provided with remedial instructions to overcome some of the problems. For example, Star and Collette (2010) report that knowing the circumstance and understanding the causes, instructor can increase the interaction with the students to provide personal interventions. As a result the students showed better academic performance and significantly increase the retention rate. In a similar study Sclater et al (2016) describe that increase interactions with students promote sense of belonging to the learner community and learning motivations. It was found that in the process the students’ attrition rate dropped from 18 to 12%. 

(ii) Supporting informed decision making 

The results from learning analytics can also be used to support informed decision making. A study by Toetenel and Rienties (2016) at the Open University in UK involves analyzing the learning designs of 157 courses taken by over 60,000 students and identify the common pedagogical patterns among the courses. The authors suggest that educators should take note of activity types and workload when designing a course and such information will be useful in decision making of specific learning design. However, the authors conclude that further studies are needed to find out whether particular learning design decisions result in better student outcomes.

(iii) Increasing cost-effectiveness 

With the funding cut and raising expenditure, cost-effective has become the key indicator for sustainability in the education sector. One of the effective ways is to take advantage of the learning management systems that not only deliver the course materials, also keep track of the learners’ activities. Instructors can analyze the activities and report the progress to the students and other stake holders in a costeffective manner. As Sclater et al (2016) note, after conducting the analysis, notifications were automatically generated and send to students and their parents on students’ performance. 

(iv) Understanding students’ learning behavior 

To better understand the students’ learning behavior, instructors can explore the data collected from the learning management systems and social media networks. Instructors can examine the relationships between students’ utilization of resources, learning patterns and preferences and learning outcomes. This approach has been adopted by Gewerc et al (2014) when attempted to examine the collaboration and social networking in a subject for education degree course. The study analyzes the intensity and relevance of the student’s contribution in the collaborative framework by using social network analysis and information extraction. The authors concluded that findings from the study help to understand more clearly how students behave during the course.

(v) Arranging personalized assistance to students 

Given the advantages of data mining techniques and algorithms that are used in business and manufacturing industry, learning analytics has emerged as educational data mining of students and the courses they study. An investigation into the application of such technique in education domain was conducted by Karkhanis and Dumbre (2015) to discover the insightful information about the students and interaction with the course. They report that after analyzing the students’ study results, demographics and social data, instructors are able to identify who need assistant most to provide individual counselling.

(vi) Providing timely feedback and intervention 

Providing feedback to students is an important role of teachers in any educational settings. This process enable students to learn from their action and can have a significant impact on motivation of the learners. The quality and timeliness of feedback are crucial in the learning process. From the learning analytics, teachers can identify students who are in need of assistance and provide appropriate intervention to the specific students. Dodge et al (2015) report that interventions through emails to the students work best and found that such approach impact on student achievement.

As the amount of data collected from the teaching-learning process increases, potential benefits of learning analytics can be far reaching to all stakeholders in education including students, teachers, leaders and policy makers. Its my firm believe that if properly leveraged, learning analytics can be an indispensable tool to narrow the achievement gap, increase student success and improve the quality of education in the digital era.

Forgetting: A Tool For Learning

Goal-Directed Forgetting

People often view forgetting as an error in an otherwise functional memory system; that is, forgetting appears to be a nuisance in our daily activities. Yet forgetting is adaptive in many circumstances. For example, if you park your car in the same lot at work each day, you must inhibit the memory of where you parked yesterday (and every day before that!) to find your car today.

Goal-directed forgetting, that is, situations in which forgetting serves some implicit or explicit personal need. In recent years research has supported the notion that mechanisms of inhibition—analogous to those proposed in many areas of lower-level cognition, such as vision (explain, perhaps parenthetically)—play an important role in goal-directed forgetting. Researchers have developed and utilized a variety of experimental paradigms to investigate phenomena that exemplify goal-directed forgetting, including directed forgetting and retrieval-induced forgetting.

Directed forgetting

Forgetting is often viewed as an uncontrollable, undesirable failure of memory. Yet it is possible to experimentally induce forgetting in an individual that can lead to unexpected benefits. One such paradigm is known as “directed forgetting.” In the typical list-based directed forgetting paradigm, a participant will study two lists of words, and is notified after each list whether or not it will be tested later on. If a list is tested after the learner was notified that it would not be tested, the learner will show weaker recall for that list, compared to a baseline condition in which all lists are expected to be tested, demonstrating the costs of directed forgetting. Interestingly, it is commonly found that recall of any list that was expected to be tested will be greater than that of the baseline condition, demonstrating the unexpected benefits of directed forgetting.

Another common paradigm for directed forgetting is the item-based method, in which participants are told after each word whether or not it will be tested. A similar pattern of results is observed, in which recall rates for the to-be-forgotten words are depressed, while recall rates for the to-be-remembered words are increased. However, the mechanisms by which item-method directed forgetting occurs are purported to be different than the mechanisms by which list-method directed forgetting operates.

In addition to studying the basic phenomenon of directed forgetting, efforts in the lab are currently underway to further investigate the effects of list-based directed forgetting using different materials and different paradigms. For example, does the pattern of results extend beyond simple word lists to more educationally relevant materials, such as text passages or videos? What happens to the pattern of results when information between the two lists is related? In addition, we are investigating whether directed forgetting applies to other learning paradigms, such as induction learning.

Retrieval-induced forgetting

Memory cues, whether categories, positions in space, scents, or the name of a place, are often linked to many items in memory. For example, the category FRUIT is linked to dozens of exemplars, such as ORANGE, BANANA, MANGO, KIWI, and so on. When forced to select from memory a single item associated to a cue (e.g., FRUIT: OR____), what happens to other items associated to that general, organizing cue? Using the retrieval-practice paradigm, we and other researchers have demonstrated that access to those associates is reduced. Retrieval-induced forgetting, or the impaired access to non-retrieved items that share a cue with retrieved items, occurs only when those associates compete during the retrieval attempt (e.g., access to BANANA is reduced because it interferes with retrieval of ORANGE, but MANGO is unaffected because it is too weak of an exemplar to interfere. Researchers argue for retrieval-induced forgetting as an example of goal-directed forgetting because it is thought to be the result of inhibitory processes that help facilitate the retrieval of the target by reducing access to competitors. In this way, retrieval induced forgetting is an adaptive aspect of a functional memory system.

In recent years, research have explored this phenomenon in a variety of ways. For example, its found that items that suffer from retrieval-induced forgetting benefit more from relearning than control items. They have also demonstrated that retrieval success is not a necessary condition for retrieval induced forgetting to occur. That is, when participants are prompted to retrieve with cues that have no possible answer (FRUIT: WO____, rather than the standard, FRUIT: OR_____), access to competing items (BANANA) is impaired, as demonstrated on a final recall test. Furthermore, researchers are currently exploring the impact of variations of the type of cue support provided for retrieval attempts (FRUIT: OR_____; FISH: ____ORE; WEAPONS: DAGG_____). Research efforts in this domain currently rest on testing various assumptions of theoretical accounts of retrieval induced forgetting.

THE ROLE OF DATA-DRIVEN FEEDBACK IN LEARNING - LEARNING ANALYTICS

Discussions about feedback frequently take place within a framing of assessment and student achievement (Black & Wiliam, 1998; Boud, 2000). In this context, the primary role of feedback is to help the student address any perceived deficits as identified through the completion of an assessment item. Ironically, assessment scores and student achievement data have also become tools for driving political priorities and agendas, and are also used as indicators in quality assurance requirements. 

Assessment in essence is a two-edged sword used to foster learning as well as a tool for measuring quality assurance and establishing competitive rankings (Wiliam, Lee, Harrison, & Black, 2004). While acknowledging the importance of assessment for quality assurance, we focus specifically on the value of feedback often associated with formative assessment or simply as a component of student completion of set learning tasks. Thus, this article explores how student trace data can be exploited to facilitate the transformation of the essence of assessment practices by focusing on feedback mechanisms. With such a purpose, I highlight and discuss current approaches to the creation and delivery of data-enhanced feedback as exemplified through the vast body of research in learning analytics and educational data mining (LA/EDM).

Although there is no unified definition of feedback in educational contexts, several comprehensive analyses of its effects on learning have been undertaken (e.g., Evans, 2013; Hattie & Timperley, 2007; Kluger & DeNisi, 1996). In sum, strong empirical evidence indicates that feedback is one of the most powerful factors influencing student learning (Hattie, 2008). The majority of studies have concluded that the provision of feedback has positive impact on academic performance. However, the overall effect size varies and, in certain cases, a negative impact has been noted. 

For instance, a meta-analysis by Kluger and DeNisi (1996) demonstrated that poorly applied feedback, characterized by an inadequate level of detail or the lack of relevance of the provided information, could have a negative effect on student performance. In this case, the authors distinguished between three levels of the locus of learner’s attention in feedback: the task, the motivation, and the meta-task level. All three are equally important and can vary gradually in focus. Additionally, Shute (2008) classified feedback in relation to its complexity, and analyzed factors affecting the provision of feedback such as its potential for negative impact, the connection with goal orientation, motivation, the presence in scaffolding mechanisms, timing, or different learner achievement levels. Shute noted that to maximize impact, any feedback provided in response to a learner’s action should be non-evaluative, supportive, timely, and specific.

Early models relating feedback to learning largely aimed to identify the types of information provided

to the student. Essentially, these studies sought to characterize the effect that different types of information can play on student learning (Kulhavy & Stock, 1989). Initial conceptualizations of feedback were driven by the differences in learning science theorisations of how the gap between the actual and desired state of the learner can be bridged (cf. historical review Kluger & DeNisi, 1996; Mory, 2004). According to Mory (2004), contemporary models build upon pre-existing paradigms by viewing feedback in the context of self-regulated learning (SRL), i.e., a style of engaging with tasks in which students exercise a suite of powerful skills (Butler & Winne, 1995). These skills, setting goals, thinking about strategies, selecting the right strategies, and monitoring the effects of these strategies on the progress towards the goals are all associated with student achievement (Butler & Winne, 1995; Pintrich, 1999; Zimmerman, 1990). As part of their theoretical synthesis between feedback and self-regulated learning, Butler and Winne (1995, p. 248) embedded two feedback loops into their model. The first loop is contained within the so-called cognitive system and refers to the capacity of individuals to monitor their internal knowledge and beliefs, goals, tactics, and strategies and change them as required by the learning scenario. The second loop occurs when the product resulting from a student engaging with a task is measured, prompting the creation of external feedback relayed back to the student; for example, an assessment score, or an instructor commenting upon the completion of a task.

Hattie and Timperley (2007) have provided one of the most influential studies on feedback and its impact on achievement. The authors’ conceptual analysis was underpinned by a definition of feedback as the information provided by an agent regarding the performance or understanding of a student. The authors proposed a model of feedback articulated around the concept that any feedback should aim to reduce the discrepancy between a student’s current understanding and their desired learning goal. As such, feedback can be framed around three questions: where am I going, how am I going, and where to next? Hattie and Timperley (2007) proposed that each of these questions should be applied to four different levels: learning task, learning process, self-regulation, and self. The learning task level refers to the elements of a simple task; for example, notifying the student if an answer is correct or incorrect. The learning process refers to general learning objectives, including various tasks at different times. The self-regulation level refers to the capacity of reflecting on the learning goals, choosing the right strategy, and monitoring the progress towards those goals. Finally, the self level refers to abstract personality traits that may not be related to the learning experience. 

The process and regulation levels are argued to be the most effective in terms of promoting deep learning and mastery of tasks. Feedback at the task-level is effective only as a supplement to the previous two levels; feedback at the self-level has been shown to be the least effective. These three questions and four levels of feedback provide the right setting to connect feedback with other aspects such as timing, positive vs. negative messages (also referred to as polarity), and the consequences of including feedback as part of an assessment instrument. These aspects have been shown to have a interdependent effect that can be positive or negative (Nicol & Macfarlane-Dick, 2006).

In reviewing established feedback models, Boud and Molloy (2013) argued that they are at times based on unrealistic assumptions about the students and the educational setting. Commonly, due to resource constraints, the proposed feedback models or at least the mechanism for generating non-evaluative, supportive, timely, and specific feedback for each student is impractical or at least not sustainable in contemporary educational scenarios. At this juncture, LA/EDM work can play a significant role in moving feedback from an irregular and unidirectional state to an active dialogue between agents.

The first initiatives using vast amounts of data to improve aspects of learning can be traced to areas such as adaptive hypermedia (Brusilovsky, 1996; Kobsa, 2007), intelligent tutoring systems (ITSs) (Corbett, Koedinger, & Anderson, 1997; Graesser, Conley, & Olney, 2012), and academic analytics (Baepler & Murdoch, 2010; Campbell, DeBlois, & Oblinger, 2007; Goldstein & Katz, 2005). Much of this research has taken place within LA/EDM research communities that share a common interest in data-intensive approaches to the research of educational setting, with the purpose of advancing educational practices (Baker & Inventado, 2014). While these communities have many similarities, there are some acknowledged differences between LA and EDM (Baker & Siemens, 2014). For example, EDM has a more reductionist focus on automated methods for discovery, as opposed to LA’s human-led explorations situated within holistic systems. Baker and Inventado (2014) noted that the main differences between LA and EDM are not so much in the preferred methodologies, but in the focus, research questions, and eventual use of models.

When considering LA/EDM through the lens of feedback, the research approaches differ in relation to the direction and recipient of feedback. For instance, LA initiatives generally provide feedback aimed towards developing the student in the learning process (e.g., self-regulation, goal setting, motivation, strategies, and tactics). In contrast, EDM initiatives tend to focus on the provision of feedback to address changes in the learning environment (e.g., providing hints that modify a task, recommending heuristics that populate the environment with the relevant resources, et cetera).

It is important to note that these generalizations are not a hard categorization between the communities, more so an observed trend in LA/EDM works that reflects their disciplinary backgrounds and interests. The following section further unpacks the work in both the EDM and LA communities related to the provision of feedback to aid student learning.

Mind: Consortium of agents

You know that everything you think and do is thought and done by you. But what's a "you"? What kinds of smaller entities cooperate inside your mind to do your work? To start to see how minds are like consortium of agents, try this:pick up a cup of tea!

Your GRASPING agents want to keep hold of the cup.

Your BALANCING agents want to keep the tea from spilling out.

Your THIRST agents want you to drink the tea.

Your MOVING agents want to get the cup to your lips.

Yet none of these consume your mind as you roam about the room talking to your friends. You scarcely think at all about Balance; Balance has no concern with Grasp; Grasp has no interest in Thirst; and Thirst is not involved with your social problems.Why not? Because they can depend on one another. If each does its own little job, the really big job will get done by all of them together: drinking tea.

How many processes are going on, to keep that tea cup level in your grasp? There must be at least a hundred of them, just to shape your wrist and palm and hand. Another thousand muscle systems must work to manage all the moving bones and joints that make your body walk around. And to keep everything in balance, each of those processes has to communicate with some of the others.What if you stumble and start to fall? Then many other processes quickly try to get things straight. Some of them are concerned with how you lean and where you place your feet. Others are occupied with what to do about the tea: you wouldn't want to burn your own hand, but neither would you want to scald someone else. You need ways to make quick decisions.

All this happens while you talk, and none of it appears to need much thought. But when you come to think of it, neither does your talk itself. What kinds of agents choose your words so that you can express the things you mean? How do those words get arranged in to phrases and sentences, each connected to the next? What agencies inside your mind keep track of all the things you've said-and, also, whom you've said them to? How foolish it can make you feel when you repeat-unless you're sure your audience is new.

We're always doing several things at once, like planning and walking and talking, and this all seems so natural that we take it for granted. But these processes actually involve more machinery than anyone can understand all at once. So, in the next few articles of this series,we'll focus on just one ordinary activity-making things with children's building-blocks. First we'll break this process into smaller parts,and then we'll see how each of them relates to all the other parts.

In doing this, we'll try to imitate how Galileo and Newton learned so much by studying the simplest kinds of pendulums and weights, mirrors and prisms. Our study of how to build with blocks will be like focusing a microscope on the simplest objects we can find, to open up a great and unexpected universe. It is the same reason why so many biologists today devote more attention to tiny germs and viruses than to magnificent lions and tigers. For me and a whole generation of students, the world of work with children's blocks has been the prism and the pendulum for studying intelligence.

In science, one can learn the most by studying what seems the least.

THE MIND AND THE BRAIN

It was never supposed [the poet Imlac said] that cogitation is

inherent in matter,or that every particle is a thinking being.Yet if

any part of matter be devoid of thought, what part can we suppose

to think? Matter can differ from matter only in form, bulk,

density, motion and direction of motion: to which of these,

however varied or combined, can consciousness annexed? To be

round or square,  to be solid or fluid, to be great or little, to be

moved slowly or swiftly one way or another, are modes of material

existence, all equally alien from the nature of cogitation. If matter

be once without thought, it can only be made to think by some new

modification, but all the modification which it can admit are

equally unconnected with cogitative powers.

                                                                                                   -Samuel Johnson

How could solid-seeming brains support such ghostly things as thoughts? This question troubled many thinkers of the past.T he world of thoughts and the world of things appeared to be too far apart to interact in any way. So long as thoughts seemed so utterly different from everything else, there seemed to be no place to start.

A few centuries ago it seemed equally impossible to explain Life, because living things appeared to be so different from anything else. Plants seemed to grow from nothing. Animals could move and learn. Both could reproduce themselves-while nothing else could do such things. But then that awesome gap began to close. Every living thing was found to be composed of smaller cells, and cells turned out to be composed of complex but comprehensible chemicals.

Soon it was found that plants did not create any substance at all but simply extracted most of their material from gases in the air. Mysteriously pulsing hearts turned out to be no more than mechanical pumps, composed of networks of muscle cells. But it was not until the present century that John von Neumann showed theoretically how cell-machines could reproduce while, almost independently, James Watson and Francis Crick discovered how each cell actually makes copies of its own hereditary code. No longer does an educated person have to seek any special,v ital force to animate each living thing.

Similarly,a century ago, we had essentially no way to start to explain how thinking works. Then psychologist like Sigmund Freud and Jean Piaget produced their theories about child development. Somewhat later, on the mechanical side,mathematicians like Kurt Godel and Alan Turing began to reveal the hitherto unknown range of what machines could be made to do. These two streams of thought began to merge only in the 1940's, when Warren McCulloch and Walter Pitt began to show how machines might be made to see, reason, and remember.

Research in the modern science of Artificial Intelligence started only in the 1950's, stimulated by the invention of modern computers. This inspired a flood of new ideas about how machines could do what only minds had done previously.

Most people still believe that no machine could ever be conscious, or feel ambition, jealousy, humor, or have any other mental life-experience. To be sure,we are still far from being able to create machines that do all the things people do. But this only means that we need better theories about how thinking works. This series of articles will show how the tiny machines that we'll call "agents of the mind" could be the long sought" particles"that those theories need.

Republic of mind: Democracy of agents run as society of plenipotentiary

   
                     Everything should be made as simple as possible, but not simpler.
                                                                                                                              -Albert Einstein

Good theories of the mind must span at least three different scales of time: slow, for the billion years in which our brains have evolved fast, for the fleeting weeks and months of infancy and childhood;    and in between, the centuries of growth of our ideas through history.

To explain the mind, we have to show how minds are built from mindless stuff, from parts that are much smaller and simpler than anything we'd consider smart. Unless we can explain the mind in terms of things that have no thoughts or feelings of their own, we'll only have gone around in a circle. But what could those simpler particles be-the "agents" that compose our minds? This is the subject of my current series of articles, and knowing this, let's see our task. There are many questions to answer.

Function: How do agents work?

Embodiment: What are they made oft

Interaction: How do they communicate?

Origins: Where do the first agents come from?

Heredity: Are we all born with the same agents?

Learning: How do we make new agents and change old ones?

Character: What are the most important kinds of agents?

Authority: What happens when agents disagree?

Intention: How could such networks want or wish?

Competence: How can groups of agents do what separate agents cannot do?

Selfness: What gives them unity or personality?

Meaning: How could they understand anything?

Sensibility: How could they have feelings and emotions?

Awareness: How could they be conscious or self-aware?

How could a theory of the mind explain so many things,when every separate question seems too hard to answer by itself? These questions all seem difficult, indeed,when we sever each one's connections to the other ones. But once we see the mind as a republic of  society of plenipotentiary run as democracy of agents, each answer will illuminate the rest. By agent in the society of plenipotentiary I mean the agent invested with full power or authority to transact business on behalf of another.

I will dig and explore further in my next article...

Skill and Skill Learning - Machine Learning Perspective

Human skill is the ability to apply past knowledge and experience in performing various given tasks. Skill can be gained incrementally through learning and practicing. To acquire, represent, model, and transfer human skill or knowledge has been a core objective for more than two decades in the fields of artificial intelligence, robotics, and intelligent control. The problem is not only important to the theory of machine intelligence, but also essential in practice for developing an intelligent robotic system. The problem of skill learning is challenging because of the lack of a suitable mathematical model to describe human skill. Consider the skill as a mapping: mapping stimuli onto responses. A human associates responses with stimuli, associates actions with scenarios, labels with patterns, effects with causes. Once a human finds a mapping, intuitively he gains a skill. Therefore, if we consider the ‘ stimuli” as input and ‘ responses” as output, the skill can be viewed as a control system. This “control system” has the following characteristics: It is nonlinear, that is, there is no linear relationship between the stimuli and responses. It is time-variant, that is, the skill depends upon the environmental conditions from time to time. 0 It is non-deterministic, that is, the skill is of inherently stochastic property, and thus it can only be measured in the statistical sense. For example, even the most skillful artist can not draw identical lines without the aid of a ruler. It is generalizable, that is, it can be generalized through a learning process. 0 It is decomposable, that is, it can be decomposed into a number of low-level subsystems. 

The challenge of skill learning depends not only upon the above mentioned inherent nature of the skill, but also upon the difficulty of understanding the learning process and transferring human skill to robots. Consider the following: A human learns his skill through an incrementally improving process. It is difficult to exactly and quantitatively describe how the information is processed and the control action is selected during such a process. 0 A human possesses a variety of sensory organs such aa eyes and ears, but a robot has limited sensors. This implies that not all human skills can be transferred to robots. The environment and sensing are subject to noises and uncertainty for a robot. These characteristics make it difficult to describe human skill by general mathematical models or traditional AI methods. 

Skill learning has been studied from different disciplines in science and engineering with different emphasis and names. The idea of learning control presented in article is based on the observation that in machine learning, actions of learning machines being subject to "playback control mode", repeat their motions over and over in cycles. The research on learning control have been reviewed anf for a repeatable task operated over a fixed duration, each time the system input and response are stored, the learning controller computes a new input in a way that guarantees that the performance error will be reduced on the next trial. Under some assumptions, the P-, PI- and PD-type learning laws have been implemented. This approach is based on control theory, but the problem is certainly beyond the domain. According to the characteristics we discussed previously, it is obviously insufficient to approach such a comprehensive problem from only a control theory point of view. The concept of task-level learning can be found in related studies. 

The basic idea is that a given task can be viewed as an input/output system driven by an input vector responding with an output vector. There is a mapping which maps task commands onto task performance. In order to select the appropriate commands to achieve a desired task performance, an inverse task mapping is needed. Task-level learning has been studied in great deal for "trajectory learning" to provide an optimum trajectory through learning and has been successful for some simple cases. For a more complicated case which is realistic in practice, the inverse task mapping is too difficult to obtain. Both learning control and task-level learning emphasize achieving a certain goal by practice, and pay no attention to modeling and learning the skill. From a different angle, a research group at MIT has been working on representing human skill. The pattern recognition method and process dynamics model method were used to represent the control behavior of human experts for a debugging process. In the pattern recognition approach, the form of IF-THEN relationship: IF(signal pattern), THEN(control action) was used to represent human skill. 

Human skill pattern model is a non-parametric model and a large database is needed to characterize the task features. The idea of the process dynamics model method is to correlate the human motion to the task process state to find out how humans change their movements and tool holding compliance in relation to the task process characteristics. The problem with this approach is that human skill can not always be represented by the explicit process dynamics model and if there is no such model, or if the model is incorrect, this method will not be feasible. Considerable research efforts have been directed toward learning control architectures using connectionist or Neural Networks. Neural Network (NN) approaches are interesting because of the learning capacity. Most of the learning methods studied by connectionists are parameter estimation methods. In order to describe the input/output behavior of a dynamic system, NN is trained using input/output data, based on the assumption that the nonlinear static map generated by NN can adequately represent the system behavior for certain applications. Although NNs have been successfully applied to various tasks, their behaviors are difficult to analyze and interpret mathematically. Usually, the performance of the NN approach is highly dependent on the architectures; however, it is hard to modify the architecture to improve the performance.

Another issue is the real-time learning, i.e., dynamically updating the model to achieve the most likely performance. In real-time circumstance, we need to compute the frequencies of occurrence of the new data and add them to the model. The procedure is the same as that used to cope with multiple independent sequences. In this study, we have shown the fundamental theory and method that are needed and the preliminary experiments for real-time learning. However, various issues on real-time learning have not been discussed extensively. For example, what happens if the measured data fed in the learning process represents the poor skill, Le., unskilled performance. Using the current method, the model will be updated to best match the performance of the operator, not to best represent the good skill. This is because we have a criterion to judge the model of the skill, but do not have a criterion to judge the skill itself. In other words, it is possible to become more unskilled in real-time learning. This is a common problem in other recognition fields such as speech recognition. One way to minimize the problem is to ensure the feeding data always represents the good performance. This again needs criterion to describe how good the skill is. We will look at this issue in the future.

In this article I presented a novel method for human skill learning using HMM. HMM is a powerful parametric model and is feasible to characterize two stochastic processes - the measurable action process and immeasurable mental states - which are involved in the skill learning. Based on “the most likely performance’! criterion, we can select the best action sequence out from all previously measured action data by modeling the skill as HMM. This selection process can be updated in real-time by feeding new action data and updating the HMM, and learning through this selection process.

The method provides a feasible way to abstract human skill as a parametric model which is easily updated by new measurement. It will be found useful in various applications in education space, besides tele-robotics, such as human action recognition in man-machine interface, coordination in anthropomorphic master robot control, feedback learning in the system with uncertainty and time-varying, and pilot skill learning for the unmanned helicopter. By selecting different units for the measured data in different problems, the basic idea is applicable for a variety of skill learning problems.