International journal of man-machine studies最新文献

It is frequently pointed out that a tabula rasa learning system needs constraints in order to extract structural information from its input-output sequence. We have been experimenting with a learning-system (PP) that incorporates a simple associative form of learning in a production system architecture. It is demonstrated that PP, implemented in a simulated robot, can learn the structure of a multi-level task with the help of speech and one or more auxiliary actions. Following a suggestion that structure could be acquired by a stress/nonstress distinction in the teacher's verbal presentation, we report briefly on an experiment that shows that stress can replace the auxiliary action.

Although the visual display unit (VDU) is becoming an increasingly popular means of displaying documents, users often show a strong preference for the "hard-copy" medium of document presentation when it comes to reading activities such as those that involve proof-reading or refereeing the document. This is partly attributed to the difficulties of annotating documents presented in the electronic medium. Voice recording may be a more acceptable medium for annotating documents that are presented on VDUs, as it overcomes many of the problems associated with the typed annotation of electronic documents. Experiment 1 compared two computer-based annotation media (typed and spoken input) with the method of writing annotations on the document. Findings suggested that writing was a superior method of annotation to the other media in terms of number of annotations elicited, speed of recording and user preference. Experiment 2 differed from the first experiment in the way in which written annotations were recorded and in the amount of pre-trial practice given to subjects. In the second experiment voice resulted in shorter task completion times than either writing or typing. This is taken as limited support for a theory that a small amount of pre-trial practice is of greater benefit to the utility of a voice annotation facility than it is to a facility for typing annotations. The majority of differences between writing and the other conditions observed in Experiment 1 were not found in Experiment 2. The contrast between the two sets of results is discussed in terms of the subjects' familiarity with the methods of annotation involved and the advantages of a facility for annotating on the document. The discussion concludes with a set of guide-lines for the implementation of a voice annotation facility.

There is a range of situations—for example, in the context of advice-giving or tutoring—in which a computer system might be required to take an active role in the interaction (rather than simply responding unquestioningly to the user's input). In such situations, the system must be able to decide how to respond to the user—sometimes taking the initiative and sometimes responding to the user's initiative. At any time, selecting the most appropriate response will depend on the context, and on what both system and user are aiming to achieve through the interaction. This paper presents the design and implementation of a computer-based agent that can engage a user in a mixed-initiative dialogue. In this work, the generation of language is viewed as opportunistic rational action. The computer-based agent constructs utterances in the context of the preceding dialogue, deciding what to say in the light of its own beliefs, goals and values. The prototype system has been tested with users. From the small-scale evaluation study that was conducted, it was concluded that the system is capable of engaging in extended dialogue that remains largely coherent and reasonable (at a semantic level), and that it provides a reasonable base for further work in this direction.

Human monitoring behavior in manual and automated scheduling systems is examined through an experiment that required the subjects to perform scheduling and monitoring tasks. The task required the assignment of a series of incoming customers to the shortest of three parallel service lines. The subject was either in charge of the customer assignment (Manual Mode) or was monitoring an automated system performing the same task (Automatic Mode). In both cases, the subjects were required to detect the nonoptimal assignments that they or the computer had made. The results showed better error detection performance and lower subjective workload in the automatic mode. The subjects in the manual mode were both biased against declaring their own assignment errors and less sensitive to their misassignments. Results are compared with previous findings of monitoring behavior in manual control systems, and are discussed in terms of human decision making, reliability, workload and system design.

A method for identifying computer users by analysing keystroking patterns with neural networks and a simple geometric distance is presented. A model of each user's normal typing style was created and compared with later typing samples. Preliminary results demonstrate complete exclusion of imposters and a reasonably low false alarm rate when the sample text was limited to the user's name.

This paper describes and discusses the design of an automatic planning system (LEAPS: the Leeds Educational Automated Planning System) for the support of educational help in information processing systems. Domain level planning is an important aspect of providing help to users of information processing systems (IPSs) as it is the mechanism that underpins answers to requests for help from the user of the form "How do I …?". Unfortunately planning technology has been inadequate for the task of supplying the answers to such problems. LEAPS provides solutions to some of the major problems of domain level planning in IPS help systems, and thus represents a technology which may be exploited to extend the flexibility and areas of application of current help systems. The difficulties of using current artificial intelligence planning technology in help systems are reviewed and the approach taken by LEAPS is described. Particular difficulties of the area, such as the absence of a complete world model and the problems of efficient but reliable plan creation, are analysed. The means by which LEAPS deals with and eases these difficulties are reviewed. LEAPS is a nearly domain-independent planner (it provides plans within a range of IPSs), which uses a constraint-based least commitment approach to the generation of its plans. The planner provides reliable plans quickly by decomposing the planning problem into its component parts and applying well understood but efficient algorithms to each component individually. The decomposed planning algorithm is reconstructed by the use of a reason maintenance system. LEAPS is able to deal with incomplete world models by applying an assumption-making plan creation algorithm, the assumptions being restricted by domain constraints representing the possible configurations of the planner's world.

Users of powerful but complex software packages do not take full advantage of the functionality of their tools. Advisory systems, or consultants, offer a solution to this problem by providing continuous and on-the-job help and training advice. However, consultants have not yet had any practical implementation outside an experimental setting. We propose an architecture for a consultant that is feasible and scalable in a practical context.

The architecture is implemented in a system called EdCoach. It addresses two important issues for advisory systems: (1) the task analysis problem and (2) the user knowledge assessment problem. The system's task analysis module infers the user's goals (task) from the analysis of actions and identifies the method chosen to complete the task. It is based on the parsing of user actions with an attribute grammar. The second component is an "overlay model" of the user's knowledge state (KS). The knowledge of the user is represented by a subset of known and unknown nodes in a set of knowledge units (KUs), representing the whole knowledge domain. The knowledge assessment module uses a probabilistic model combined with an implication network to infer user knowledge from the result of the task analysis.

A third component of the system is the didactic module, which consists essentially in the application of a straightforward principle: if the user adopts an inefficient method to complete a goal, the system first checks that the efficient method is unknown and, if so, it advises the user about that method.

The system's performance was empirically tested with a text-editing application. A simulation of all three modules integrated in EdCoach shows that after about two weeks, 75% of the potential recommendations were progressively and correctly administered, or withheld, according to whether or not the efficient methods were known or unknown. The advantages and limits of the general approach adopted in EdCoach are discussed.

Semiotic approaches to design have recently shown that systems are messages sent from designers so users. In this paper we examine the nature of such messages and show that systems are messages that can send and receive other messages—they are metacommunication artefacts that should be engineered according to explicit semiotic principles. User interface languages are the primary expressive resource for such complex communication environments. Existing cognitively-based research has provided results which set the target interface designers should hit, but little is said about how to make successful decisions during the process of design itself. In an attempt to give theoretical support to the elaboration of user interface languages, we explore Eco's Theory of Sign Production (U. Eco, A Theory of Semiotics, Bloomington, IN: Indiana University Press, 1976) and build a semiotic framework within which many design issues can be explained and predicted.

This paper presents five abstract characteristics of the mental representation of computer programs: hierarchical structure, explicit mapping of code to goals, foundation on recognition of recurring patterns, connection of knowledge, and grounding in the program text. An experiment is reported in which expert and novice programmers studied a Pascal program for comprehension and then answered a series of questions about it, designed to show these characteristics if they existed in the mental representations formed. Evidence for all of the abstract characteristics was found in the mental representations of expert programmers. Novices' representations generally lacked the characteristics, but there was evidence that they had the beginnings, although poorly developed, of such characteristics.

The Dempster Sharer theory of evidence concerns the elicitation and manipulation of degrees of belief rendered by multiple sources of evidence to a common set of propositions. Information indexing and retrieval applications use a variety of quantitative means—both probabilistic and quasi-probabilistic—to represent and manipulate relevance numbers and index vectors. Recently, several proposals were made to use the Dempster Shafer model as a relevance calculus in such applications. This paper provides a critical review of these proposals, pointing at several theoretical caveats and suggesting ways to resolve them. The methodology is based on expounding a canonical indexing model whose relevance measures and combinations mechanisms are shown to be isomorphic to Shafer's belief functions and to Dempster's rule, respectively. Hence, the paper has two objectives: (i) to describe and resolve some caveats in the way the Dempster Shafer theory is applied to information indexing and retrieval, and (ii) to provide an intuitive interpretation of the Dempster Shafer theory, as it unfolds in the simple context of a canonical indexing model.