Thesis (M.S.) -- Worcester Polytechnic Institute. / Keywords: Intelligent tutoring systems; Behavior; Learning; Visualization. Includes bibliographical references (leaves 37-38).

This thesis examines children's arithmetic strategies and their relation to the concepts of commutativity and associativity. Two complementary methods were used in this research: empirical studies and computational models. Empirical studies were carried out to identify the strategies children used for solving problems like 3 + 4, and 3 + 4 + 7, and the conceptual knowledge associated with them. Their understanding of subtraction problems where the minuend is less than the subtrahend (e.g. 6-8) was also considered. A study with 105 subjects revealed a variety of strategies and information about children's knowledge of commutativity and associativity. Four levels of performance of commutativity were also identified. A longitudinal study was carried out with 12 children in order to obtain details of children's changes in strategy, and to double check the results obtained in the main study. The strategies observed to be used by children over a 20 month period parallel those found in previous studies, which show a general transition to more efficient methods. However, the longitudinal study revealed that development of such arithmetic strategies is a slow process. Furthermore, the studies indicated that knowledge of commutativity is a prerequisite for associativity. Models of the observed strategies have been implemented in the form of production rules in a computer program called PALM. The process of implementation highlighted features of children's problem solving that had not been 'detected during the studies. In addition to models that describe the space of strategies, a model of learning has been implemented for the transition from procedural knowledge of commutativity to that of associativity. The model is capable of generalizing its inbuilt knowledge, for instance, its ability to solve 2-term arithmetic expressions, to allow it to solve more complex problems, such as 3-term arithmetic expressions. A further model has been constructed for learning arithmetic strategies that are more efficient than those already represented in the program. It learns specific rules by adding conditions for efficient problem solving to its previous general rules.

As the range of models which tutoring systems can capture is extended, efficient diagnosis becomes more difficult. This thesis describes a solution to this problem based on the generation of 'Critical Problems'; their role in student modelling is analogous to that of the 'Crucial Experiment' in science. We argue that great diagnostic power can be obtained by generating discriminatory problem examples. In general, efficient diagnosis is just not possible without such an hypothesis-testing capability. We describe a program, PO, which given a pair of production rule models and a description of the class of problems which the student must solve, generates an abstract specification of the problems which discriminate between those two hypotheses. Through a process termed 'Abstract Interpretation', PO tips the balance in favour of diagnostic measurement. The key to this problem lies in the realisation that we are only interested in the abstract mapping between a model's inputs and outputs; from the point of view of generating a Critical Problem, the intermediate processing of the model is irrelevant.

In the history of Intelligent Tutoring Systems, SOPHIE (Brown, Burton, and Bell, 1974), now considered a classic, contained many important ideas and features. One of these was its natural language user interface. Today, the trend has moved away from natural language interfaces towards graphical ones although the argument in favour of natural language user interfaces, both from Human Computer Interaction and natural language researchers, still persist. Is this argument correct? This thesis explores this question by investigating how SOPHIE might be re-implemented with a graphical direct manipulation interface instead of a natural language one, with the goal of improving its standard of usability. It begins by analysing the features that seem to have been central to SOPHIE's usability. These, it argues, were not so much an ability to accept well formed complete English sentences, as an ability to accept and interpret correctly a wide range of abbreviated inputs. Two models of interaction, Circuit 1, a pilot, and Circuit II, a fairly full implementation of SOPHIE were implemented and tested. Both employ free-order syntax that allows users to specify the components of a full command in any order. The combination of deixis and free-order syntax supported allows completely general ellipsis which achieves, in extended interaction sequences, the same economy and naturalness that SOPHIE achieved through its use of anaphora and ellipsis. Whilst the free-order syntax. technique is little used at present in user interfaces, the results of observational studies conducted have shown that it saves users time and convenience. Thus, considering key linguistic features of a natural language user interface has shown how novel features can enhance the usability of direct manipulation interfaces. This thesis argues that user interfaces can be improved by employing structures found in natural language or at least conversation which can be constructed within direct manipulation interface styles. This approach was further expanded to support topic shifts between different circuit contexts. Circuit II, like SOPHIE, supports three different topics: normal circuit behaviour, a circuit with an unknown fault, and circuits with user-hypothesised faults. Drawing on Reichman's (1981) work, Circuit II uses natural language cue phrases of the type "by the way", re-implemented in the direct manipulation style, to facilitate shifts between topics in a smoother and more natural way than SOPHIE which , used clumsy explicit commands.

No description available.

No description available.

One of the greatest problems facing researchers in the sub field of Artificial Intelligence known as Intelligent Tutoring Systems (ITS) is the selection of a knowledge base designs that will facilitate the modification of the knowledge base. The Class-Entity-Relationship-Attribute (CERA), proposed by R. P. Brazile, holds certain promise as a more generic knowledge base design framework upon which can be built robust and efficient ITS. This study has a twofold purpose. The first is to demonstrate that a CERA knowledge base can be constructed for an ITS on a subset of the domain of Cretaceous paleontology and function as the "expert module" of the ITS. The second is to test the validity of the ideas that students guided through a lesson learn more factual knowledge, while those who explore the knowledge base that underlies the lesson through query at their own pace will be able to formulate their own integrative knowledge from the knowledge gained in their explorations and spend more time on the system. This study concludes that a CERA-based system can be constructed as an effective teaching tool. However, while an ITS - treatment provides for statistically significant gains in achievement test scores, the type of treatment seems not to matter as much as time spent on task. This would seem to indicate that a query-based system which allows the user to progress at their own pace would be a better type of system for the presentation of material due to the greater amount of on-line computer time exhibited by the users.

This dissertation consists of two articles. The first article describes an architecture for intelligent tutoring that focuses on modularity. This new architecture is based on Gibbons' layers theory for instructional design (2014). Splitting up the architecture for an intelligent tutor into layers allows different pieces to age at different rates which, in turn, allows the intelligent tutor to be adapted to new research and design theories. This architecture supports building intelligent tutoring services, nimble programs that can be assembled together to replicate the functions of intelligent tutoring without the expertise needed to create the services. Alternative architectures support building intelligent tutoring systems, monolithic programs that are less amenable to change and require immense expertise. The second article provides a proof of concept for the first services created under the layers theory. These two services create the building blocks of a domain and comprise one part of the content layer as described in the first article. The first service focuses on the task of key concept extraction whereas the second service focuses on prerequisite relationship extraction. These two tasks can provide the structure of the domain, particularly when it comes to domains that are more declarative in nature rather than procedural.

This paper sought to accomplish three goals. First, it provided a systematic, comparative review of several intelligent tutoring systems (ITS). Second, it summarized problems and solutions presented and solved by developers of ITS by consolidating the knowledge of the field into a single review. Third, it provided a unified language from which ITS can be reviewed and understood in the same context. The findings of this review centered on the 5-Component Framework. The first component, the domain model, showed that most ITS are focused on science, technology, and mathematics. Within these fields, ITS generally have mastery learning as the desired level of understanding. The second component, the tutor model, showed that constructivism is the theoretical strategy that informs most ITS. The tutoring tactics employed in the ITS stem from this paradigm. The third component, the student model, describes the several ways ITS infer what a student knows. It described the variety of data that is collected by an ITS and how it is used to build the student model. The fourth component, the interface, revealed that most ITS are now web-based, but vary in their capacity to interact with students. It also showed that user experience is underreported and ought to be included more in the research. Finally, the fifth component, learning gains, demonstrated that ITS are capable of producing learning gains equivalent to a human tutor. However, reporting learning gains does not seem to be a focus of the literature.

With the help of the internet, teaching is not constrained in the traditional classroom pedagogy; the instructors can put the course material on the website and allow the students go on to the course webpage as an alternative way to learn the domain knowledge. The problem here is how to design a web-based system that is intelligent and adaptive enough to teach the students domain knowledge in Programmable Logic Controller (PLC). In my research, I proposed a system architecture which combines the pre-test, cased-based reasoning (i.e., heuristic functions), tutorials and tests of the domain concepts, and post-test (i.e., including pre-exam and post-exam) to customize students’ needs according to their knowledge levels and help them learn the PLC concepts effectively. I have developed an intelligent tutoring system which is mainly based on the feedback and learning preference of the users’ questionnaires. It includes many pictures, colorful diagrams, and interesting animations (i.e., switch control of the user’s rung configuration) to attract the users’ attention. From the model simulation results, a knowledge proficiency effect occurs on problem-solving time. If the students are more knowledgeable about PLC concepts, they will take less time to complete problems than those who are not as proficient. Additionally, from the system experiments, the results indicate that the learning algorithm in this system is robust enough to pinpoint the most accurate error pattern (i.e., almost 90 percent accuracy of mapping to the most similar error pattern), and the adaptive system will have a higher accuracy of discerning the error patterns which are close to the answers of the PLC problems when the databases have more built-in error patterns. The participant evaluation indicates that after using this system, the users will learn how to solve the problems and have a much better performance than before. After evaluating the tutoring system, we also ask the participants to submit the survey (feedback), which will be taken into serious consideration in our future work.