In contrast to current intelligent systems, which must be laboriously programmed for each task they are meant to perform, instructable agents can be taught new tasks and associated knowledge. This thesis presents a general theory of learning from tutorial instruction and its use to produce an instructable agent. Tutorial instruction is a particularly powerful form of instruction, because it allows the instructor to communicate whatever kind of knowledge a student needs at whatever point it is needed. To exploit this broad flexibility, however, a tutorable agent must support a full range of interaction with its instructor to learn a full range of knowledge. Thus, unlike most machine learning tasks, which target deep learning of a single kind of knowledge from a single kind of input, tutorability requires a breadth of learning from a broad range of instructional interactions. The theory of learning from tutorial...

Abstract not found

Modern computer systems routinely present information to the user as a combination of text and diagrammatic images, described as "graphical user interfaces". Practitioners and researchers in Human-Computer Interaction (HCI) generally believe that the value of these diagrammatic representations is derived from metaphorical reasoning; they communicate abstract information by depicting a physical situation from which the abstractions can be inferred. This assumption has been prevalent in HCI research for over 20 years, but has seldom been tested experimentally. This thesis analyses the reasons why diagrams are believed to assist with abstract reasoning. It then presents the results of a series of experiments testing the contribution of metaphor to comprehension, problem solving, explanation and memory tasks carried out using a range of different diagrams. The results indicate that explicit metaphors provide surprisingly little benefit for cognitive tasks using diagrams as an external re...

Up until now there has been no agreement on what a mental model of a physical system is and how to infer the mental model a person has. This paper describes research aimed at solving these problems by proposing that a Mental Model is a dynamic representation created in WM by combining information stored in LTM (the Conceptual Model of the system) and characteristics extracted from the environment. Three experiments tested hypotheses derived from this proposal. Implications for research on Mental Model are discussed. Mental Models and Working Memory 3 3 The role of Working Memory on measuring Mental Models of physical systems When a person learns to interact with a system it means she/he acquires knowledge about its operation and about the structural relationships between its components. Researchers have called this knowledge the 'Mental Model' of the system (Moran, 1981). The existence of Mental Models, and their importance during the interaction with the system, has been demonstr...

Abstract—The nature of an information visualization can be considered to lie in the visual metaphors it uses to structure information. The process of understanding a visualization therefore involves an interaction between these external visual metaphors and the user’s internal knowledge representations. To investigate this claim, we conducted an experiment to test the effects of visual and verbal metaphor on the understanding of tree visualizations. Participants answered simple data comprehension questions while viewing either a treemap or a node-link diagram. Questions were worded to reflect a verbal metaphor that was either compatible or incompatible with the visualization a participant was using. The results (based on correctness and response time) suggest that the visual metaphor indeed affects how a user derives information from a visualization. Additionally, we found that the degree to which a user is affected by the metaphor is strongly correlated with the user’s ability to answer task questions correctly. These findings are a first step towards illuminating how visual metaphors shape user understanding and have significant implications for the evaluation, application, and theory of visualization. Index Terms—Cognition, visualization theory, metaphors, hierarchies, evaluation. 1

The authors introduce a model of skill acquisition that incorporates elements of both traditional models and models based on embedded cognition by striking a balance between top-down and bottom-up control. A knowledge representation is used in which pre- and postconditions are attached to actions. This model captures improved performance due to learning not only in terms of shorter solution times and lower error rates during the task but also in an increased flexibility to solve similar problems and robustness against unexpected events. In 3 experiments using a complex aviation task, the authors contrasted instructions that explicitly stated pre- and postconditions with conventional instructions that did not. The instructions with pre- and postconditions led to better and more robust performance than other instructions, especially on problems that required transfer. The parameters of the model were estimated to obtain a quantitative fit of the results of Experiment 1, which was then successfully used to predict the results of Experiments 2 and 3.

Intelligent Student Systems are a class of Intelligent Learning Environments that place the learner in the role of a tutor rather than a student. In an analogy with the educational practice of peer tutoring users learn by teaching the computer -- inverting the predominant `computer as tutor' metaphor. Intelligent Student Systems emphasize the learner's viewpoint in educational interactions in preference to the system's conception of the domain. These systems are considered to be less complex than Intelligent Tutoring Systems and to have the potential to generate novel human-computer educational interactions. Viewpoints also have an integral part in knowledge representation in Intelligent Learning Environments and they are utilised in the design and implementation of an Intelligent Student System in economics. Testing of the system produced insights into the future application of Intelligent Student Systems.

We have developed a process model that learns in multiple ways using the Soar chunking mechanism while finding faults in a simple control panel device. The model accounts very well for measures such as problem solving strategy, the relative difficulty of faults, average fault-finding time, and, because the model learns as well, the speed up due to learning when examined across subjects, faults, and even series of trials for individuals. However, subjects tended to take longer than predicted to find a fault the second time they completed a task. To examine this effect, we compared the model's sequential predictions -- the order and relative speed that it examined interface objects -- with a subject's performance. We found that (a) the model's operators and subject's actions were applied in basically the same order; (b) during the initial learning phase there was greater variation in the time taken to apply operators than the model predicted; (c) the subject appeared to spend time checking their work after completing the task (which the model did not). The failure to match times on the second time seeing a fault may be accounted for by the subject spent checking their work whilst they learn to solve the fault-finding problems. The sequential analysis reminds us that though aggregate measures can be well matched by a model, the underlying processes that generate these predictions can differ.

The software interface C whether graphical, command-oriented, menu-driven, or in the form of subroutine calls C shapes the user's perception of what software can do. It also establishes upper bounds on software usability. Numerical software interfaces typically are based on the designer's understanding of how the software should be used. That is a poor foundation for usability, since the features that are "instinctively right " from the developer's perspective are often the very ones that technical programmers find most objectionable or most difficult to learn. This paper discusses how numerical software interfaces can be improved by involving users more actively in design, a process known as user-centered design (UCD). While UCD requires extra organization and effort, it results in much higher levels of usability and can actually reduce software costs. This is true not just for graphical user interfaces, but for all software interfaces. Examples show how UCD improved the usability of a subroutine library, a command language, and an invocation interface. Improving the Usability of Numerical Software through User-Centered Design A "build it and they will come " mentality has dominated the design of scientific software for some time. It is increasingly clear, however, that this attitude is responsible for the failure of many software systems. Software users are no longer willing to put up with products that are difficult to learn or use [7].

We have developed a process model of problem solving with a simple control panel device. The model accounts well for many aggregate measures, including those from a study reported here (N=10): problem solving strategy, average fault-finding time, and the relative difficulty of faults. To further test the model, we compared the model's sequential predictions -- the order and relative speed that it examined interface objects and answered with a subject solving five tasks, making it one of the first models to have its sequential predictions compared with human data as they both learn. We found that the predicted actions matched and mismatch in systematic ways. The correspondences showed that: (a) subjects were reflecting or checking their work; suggesting that learning mechanisms can, in some instances, model the speed of learning while not completely modelling the mechanism; (b) mouse movements in the interface provide support for the model's predictions of what the subject attended to; providing further evidence that protocols can be augmented with mouse traces; and (c) while the comparison of sequential predictions may be becoming more tractable, the effects of learning raises new difficulties, such as assigning credit to model structures that change with performance.