Chapter 2 Learning by Computer
2.5 Intelligent Tutoring Systems
2.6.4 Facilitating learning
Before assessing the effectiveness of teaching methods one needs to define what is meant by the term 'instruction' and to have a theory of instruction. Resnick ( 1 985a; 1 9 85b) maintains that instruction is any activity that facilitates the process of learning in another person. That is, it is an intervention procedure in a learner's ongoing processes of knowledge acquisition. She postulates that a theory of instruction should have three components: a theory of intervention (how the instructor operates), a theory of acquisition as described in the previous section, and a theory of expertise (some criteria for measuring the student's competence level in the domain). Resnick believes that theories of expertise are well advanced but that acquisition is not so well understood. Even less is known about effective methods of intervention.
She looks at various aspects of the problem of appropriate intervention, observing that we can no longer think of teaching as a process of directly putting knowledge or skill into people. Instead 'effective instruction must aim to place learners in situations where the constructions that they naturally make as they think about the events and information that impinge on them are maximally likely to be correct and efficient' (Resnick, 1 9 85b, pp28- 29). Also, the instructor should not necessarily be looking for ways of presenting information that directly match the thought or performance patterns of experts. Rather s/he should be trying to find instructional representations that allow learners to construct those expert representations gradually for themselves.
Glaser and his colleagues (Glaser et al. , 1 985) come to a similar conclusion when testing avionics technicians. The technicians had previously been taught by demonstration of the algorithms, but, since they had not devised these themselves, they learned nothing of the diagnostic and procedural skills that are needed. The students needed to apply problem solving in the specialty in order to develop stronger problem solving methods from the known weak ones. The LRDC team recommend that the students be supplied with more declarative knowledge of the specialty, rather than having to determine it, and be given more opportunity to practise problem solving. They propose a simulation method whereby the technicians may develop these skills, improving their competence by tackling and solving novel problems. They observe that experts in a routine domain often do not handle novel problems well.
One of the important characteristics of effective problem solving and of learning in the process seems to be self-monitoring. Glaser and his colleagues (Glaser et al. , 1 985) discuss meta-cogniti�e strategies or comprehension monitoring (ie establishing learning goals, assessing the degree to which these goals are being met, and modifying strategies
and methods when necessary to better achieve these goals). Glaser ( 1 989) suggests that the ability to self-monitor (for example, to assess how far through the solution process one has progressed and how much work each part of the problem will take) is a measure of expertise.
Res nick ( 1 985b) observes the same kind of phenomenon. Some students automatically monitor their own progress and adjust their activities accordingly. For example, she notes that students will often re-read a piece of text that they could not make sense of first time through. Further improvement can be achieved by encouraging the subjects to suggest questions about a piece of text that could be asked of others. Also they might be asked to predict what happens next.
Weinstein and Rogers ( 1 984) at BBN find that comprehension monitoring can be quickly and effectively taught by encouraging students to engage in periodic self-questioning when studying textbook materials. One approach involves the tutor and student taking turns in p araphrasing, summarizing and asking appropriate questions on alternate paragraphs of a piece of text: using a so-called Reciprocal Teaching Method.
It is suggested by one group (Chipman et al. , 1 985) that it is difficult to teach general problem solving skills to someone who does not have knowledge of a specific domain. This confirms the findings of the LRDC team (Glaser et al. , 1985) that specific problem solving skills seem to be the main discriminant between the more and less skilled workers. Seeming to contradict this is the notion that specific problem solving comes from applying general problem solving with the special knowledge of the domain. Glaser and his colleagues recognise this apparent contradiction and suggest that the two should be developed in parallel with a specific domain but, once this is done, the techniques developed can be applied in other domains: 'An intelligent practice environment should provide considerable external support in the form of general prompts for planning, monitoring, and modelling, for general problem solving skills which may not be developed well enough to be exercised autonomously . . . The goal of intelligent training systems should be to provide practice in applying general problem solving skills to specific problems that are of the same kinds as the trainee will face on the job. This practice then will allow specialized, expert forms of the higher-order skills to be proceduralized and strengthened' (Glaser et al. , 1 985, p292). Although this means that the subject will have only specific knowledge of a single domain, it is believed that slhe will have learned enough to be able to adapt quickly to related ones.
The importance of an appropriate model of the domain is also mentioned by Chi (LRDC, 1 986). She suggests that, instead of trying to teach children better ways of accessing
their knowledge, we should be ensuring that they build up better internal knowledge structures. If the student has an inappropriate model then misconceptions can result (Resnick, 1 985b). The instructor should try to ensure that the subject does not develop 'buggy' models. In cases where the subject has developed obvious misconceptions the instructor must decide whether to confront the subject directly with the new model or let the student discover the inconsistencies him/herself.
Champagne, Klopfer and Gunstone ( 1 982) find one effective way of reducing misconceptions is to encourage students to discuss their beliefs among themselves and to adopt an i nitially more qualitative view of what they are learning than is normally taken. In later papers (Champagne et al. , 1 985b; Champagne et al., 1 9 85 a) this approach is formalised as an 'ideational confrontational strategy'.
Minimising the chances of misconceptions can be achieved by proceeding with caution. Learning is facilitated by presenting subjects with problems that they should be able to solve with their current knowledge but which extend them and encourage them to modify their knowledge structures. This is particularly effective if the subject has a large declarative knowledge of the domain and discovery techniques are used for exploration (Glaser, 1 987).
Misconceptions can often arise when inappropriate heuristics have been taught. Often heuristics are used to aid the learning process of novices. This is understandable since these rules of thumb provide shortcuts and can be applied without requiring a detailed knowledge of the domain. Glaser and Chi ( 1 988), however, observe that inappropriate heuristics can, in fact, hamper novices' development of expertise since the rules often have no logical basis, the student may become too dependent upon them, and they may not work with a more elaborate model of the domain.
Short term memory capacity seems to be another limiting factor on learning (Resnick, 1 985b ), consequently the tutor should try to reduce the demands on memory. This can be done by encouraging automation of simpler tasks and by helping the subject to develop appropriate groupings into chunks. Resnick believes that the development of understanding is affected by the kinds of practice afforded by instruction and by the ways in which procedural practice is interspersed with invitations to reflect and construct explanations. Shute, Glaser and Raghavan ( 1 989) also maintain that understanding is not necessarily a pre-cursor to performance. Often it follows on later from the ability to solve problems. Procedural knowledge within the domain is more crucial than understanding. In particular, subjects learn well in a discovery environment where they can formulate hypotheses and then test them.
The use of a discovery environment that students can explore is examined by Shute and her colleagues (Shute et al. , 1 989). By proposing and testing hypotheses the student learns about the environment. If the learning is done by getting the subject to solve problems then these problems should be carefully chosen to extend the subject's knowledge gradually so that s/he can build on existing knowledge.
As noted in Section 2.6.2, experts seem to use different (and fewer) weak problem solving methods from those of novices. Can expertise be improved, then, by getting novices to adopt the weak problem solving methods used by experts? The answer seems to be no. The distinction between weak problem solving methods used by experts and novices would appear to be an effect rather than a cause. For example, self-regulation helps in solving problems but only if one has enough knowledge of problem solving methods within the domain to know how well they are working. Rabinowitz and Glaser ( 1 986) observe that although experts tend to use a working forward strategy, this would not necessarily help novices since they do not have the domain knowledge to exploit such a technique. Gobbo and Chi ( 1 986) find that inferencing is an effective technique used by experts for problem solving but that the amount of inferencing that can be done is very much dependent on their knowledge of the domain. Thus, it is because the expert has a well-structured knowledge base, slhe can effectively apply certain weak problem solving methods. It may not be appropriate to attempt to teach novices similar methods.
Although strong methods appear to be ultimately more important than weak ones i n solving problems in particular specialties, there are important inter-dependencies between them. For example, Glaser's team (Glaser et al. , 1 985) suggest that strong methods are developed from the operation of weak methods on declarative knowledge. It is also clear that without specific domain knowledge many weak methods would be ineffective.