Read an Excerpt
Chapter 1: Introduction
Teaching of knowledge requires a detailed and very explicit model of the domain or subject matter. Since teaching usually involves more than just the presentation of correct knowledge, identification and diagnosis of deviations in the learner's knowledge is needed. This requires subject matter modelling to be even more extensive. One of the major bottlenecks in developing educational systems is the fact that the construction of subject matter models is a laborious task, while the result is mostly not reusable. The subject matter representation is often specific to the task to be taught, and diagnosis of the learner's errors is mostly based on explicit fault models or bug catalogues. In this book we present a representation and a set of algorithms that facilitate automated generation of subject matter models representing qualitative system behaviour. Furthermore, diagnostic techniques are presented that exploit these subject matter models and for which no explicit fault models are required.
The research reported here focusses on teaching to understand the behaviour of systems by means of practice. In every-day life, people have to interact with and reason about a large number of systems. This includes physical devices such as door knobs, taps, and stereo equipment, as well as non-physical systems when dealing with, for instance, stock markets or weather forecasting. Also in professional life a growing number of people has to be trained in operating complex devices such as airplanes or power plants. The teaching goal may vary from inducing insight in the physical principles underlying the behaviour of the device to teaching behaviour analysis in the context of system operation and maintenance. These goals strongly overlap, but the former goal is typically emphasised in high school education and the latter in professional training.
Computer simulations can provide interactive environments by means of which people can develop knowledge about the behaviour of complex systems. The steady increase in computing power has given simulation a solid position within the area of educational systems (de Jong & Sarti, 1994). A possible pitfall when using simulation in education is what has been called the 'science museum problem' (Twidale, 1993). The metaphor expresses the stereotype behaviour of learners to run around pulling handles and pressing buttons, without actually understanding what is \going on inside the devices they are manipulating. Several studies have shown that learners easily 'get lost' in a simulation environment if no proper guidance is provided (e.g., Elsom-Cook, 1990; Kamsteeg, 1994; van der Hulst, 1996). To provide guidance for the teaching goals mentioned above, the educational system should be able to communicate about the simulation itself. To this end, the simulation model needs to be articulate (Forbus, 1990; Forbus, 1991): the model should be rich enough and contain the right handles to facilitate the communication about its contents.
Qualitative simulators, such as QPE (Forbus, 1990) and GARP (Bredeweg, 1992), provide an excellent basis for articulate models. This is not really surprising, because the field of qualitative reasoning was initiated to overcome the communicative shortcomings of quantitative simulators in teaching systems such as STEAMER (Hollan et al., 1984) and SOPHIE I (Brown et al., 1982). What is surprising, is that qualitative reasoning has rapidly become a separate field. Relatively few research projects have dealt with applying qualitative models in teaching systems (SOPHIE I I I (Brown et al., 1982), QUEST (White & Frederiksen, 1990), ITSIE (Sime & Leitch, 1992), MACH-111 (Massey et al., 1988), and MULEDS (Plotzner & Spada, 1992)). There are even less educational systems that actually incorporate a qualitative simulator. Forbus' work on Self-Explanatory Simulators (Forbus, 1996; Forbus, 1997) is one of the few examples, although this is not a purely qualitative simulation.
The research presented aims at exploiting the representation and reasoning capacities of qualitative reasoning techniques in supporting the interaction with a learner.
1.1 Individualised Interaction
We consider interaction to be a key aspect in learning. The educational context may vary from young children learning basic qualitative principles of physics by interacting with a simulation, to industrial personnel that are trained to operate or maintain a complex device. In each of these situations, learning takes place by interacting with this learning context. To avoid problems like the 'science museum problem' mentioned earlier, knowledgeable guidance is a necessary condition for successful learning.
To facilitate knowledgeable guidance in computer-based learning environments, these systems should fulfill two requirements. Firstly, the educational system should know about the subject matter it tries to communicate. That is, it should contain an articulate model of this subject matter. Secondly, the system should be able to adapt the interaction to the situation at hand; not only to the specific subject that is considered, but also to the individual learner it is interacting with. This individualisation requires knowledge about the learner, and hence involves assessment of his or her knowledge: the educational system has to diagnose the learner's problem solving behaviour. The role of diagnosis is not limited to once in a while determining some specific error made by the learner. Instead, the diagnoser plays a central role in guiding the interaction.
1.2 Diagnosis of Problem Solving Behaviour
The teaching goal we focus on can be more precisely defined as the acquisition of problem solving tasks such as predicting or 'postdicting' (explaining) the behaviour of a device (cf Breuker, 1994a). Hence, the learner's behaviour consists of a set of inferences about the behaviour of this device.' In order to guide the learner in an adaptive way, interpretation and diagnosis of this set of inferences is required. Because of the complexity of problem solving tasks, not all individual inferences involved in the process can be reported by the learner; this would result in numerous reiterations of details and would soon reduce the motivation of the learner. The approach we take is to monitor the problem solving behaviour of the learner at a global level. When the learner's behaviour deviates from the behaviour predicted by the subject matter model, diagnosis is needed to identify in more detail where this deviation is located in the model. Hence, diagnosis is performed as soon as the educational system needs to know about inference behaviour that has not been explicitly reported by the learner. To facilitate this, the subject matter model should contain all the individual inference steps needed for the problem solution. When the educational situation requires so, this knowledge is available for broadening or deepening the interaction. The role of diagnosis is exactly to determine which part of this knowledge should be selected.
In research on artificial intelligence in education, individualisation and diagnosis are generally considered extremely difficult, or even infeasible. One important reason is that behaviour diagnosis' is usually based on hand-crafted catalogues of misconceptions, or bugs, which makes diagnostic systems very expensive to develop. Because there can be infinitely more erroneous than correct responses, development of bug catalogues requires systematic empirical studies. Moreover, a catalogue of bugs is only applicable to one specific domain. When the subject matter changes, a new catalogue has to be developed and implemented. We therefore need generic methods instead of case-specific knowledge about errors and misconceptions.
In order to ensure the generality of the diagnostic method, we propose a model-based diagnosis approach (cf Hamscher et al., 1992b). This choice is based on the fact that modelbased diagnosis is an intensively studied and well-understood field of research, and hence we can reuse existing techniques and representations (Self, 1993). Moreover, model-based diagnosis does not require explicit fault models or bug catalogues for its operation, but instead reasons from a representation of the correct behaviour of the system to be diagnosed.
The same diagnostic component can be applied to different subject matter models, as long as they comply with the modelling principles required by the diagnostic algorithms. The main requirement is that the system under diagnosis can be modelled as a set of connected components, for which the behaviour can be defined individually. These components should model the smallest entities that can be individually repaired; diagnosis is only useful down to the level of possible repairs. In the context of learning, this means that these components should represent the smallest units of problem solving knowledge that are still relevant in an educational setting (e.g., that can be individually explained). The constraints that the component-connection paradigm puts on the subject matter models are not only a requirement for diagnosis, but are also useful for other educational functions: a clear separation of the knowledge in terms of indexed elements that make sense in education is a prerequisite for automated structured explanation, generation of questions, or subject matter sequencing. Only with such a modular structure, different components can be freely combined, selected, and sequenced...
loading...