This research has produced different types of outcomes both at the design level and at the implementation level. Below, we list both the outcomes of this research project and some avenues for continuing research in this domain:
- We developed an original solution for separating domain-independent from domain-specific knowledge. Instead of searching for a domain-independent language able to describe any domain, we describe pedagogical knowledge in terms of interactions among agents. A learning environment is a society of agents for which we have specified (1) the role distribution over agents (coach, tutor, experts) and (2) protocols of communication among agents (global and local interaction mode, repair, conflicts,...). Some of the choices made in ETOILE are still arbitrary. For instance, the only domain-specific agents we implemented are `experts' while we could integrate `peer learners' or semi-experts. Another limitations is that `agents' refer to a simple data-structure, including namely a rulebase that drives his behavior. Current agents do not have the autonomy and goal-orientedness that characterize a proper multi-agent system (Soham, 1993). The specification of well-defined communication protocols among agents would enable the integration of heterogeneous agents, namely agents which do not rely on rule bases or agents written in a different programming language. This approach would also extend ETOILE's scope to designers outside the LISP community.
- By separating pedagogical from domain knowledge, we are not only capable of permuting the domain knowledge associated with some teaching strategy, but conversely, we are also able to apply various teaching styles to a particular domain expertise. Instead of arguing which teaching style is the best, we prefer to cover a large range of teaching styles. Of course, this raises the issue of selecting the teaching style. Despite this issue being non-trivial from some theoretical viewpoint, we solved it with a simple pragmatic principle: if the learner fails, change the teaching style! We would like to develop future research on this learner's perception of teaching styles.
- A third outcome of our work lies in the design metaphors used for developing MEMOLAB and ETOILE. The pyramid metaphor is in itself very simple, but it enabled us to structure our sequence of microworlds on the basis of some theory of cognitive development. This metaphor is closely connected to the language shift principle which helps developers to smooth the transition between two microworlds using different interaction languages. In the near future, we would like to collect more empirical data to check the theory's validity and its application to MEMOLAB, and to observe the learner adaptation mechanisms when a language shift occurs.
- Another achievement concerns the integration of multiple tools within a learning environment, especially the integration of a hypertext with rulebase agents. Nowadays, hypertext systems are a very popular research topic. We believe that the pedagogical power of a hypertext cannot be understood without considering the context that justifies information search in a hypertext.
- A fifth lesson from this work concerns the interaction between the leaner and a computerized agent. The principle we applied is very simple: human-computer interaction should concern what is ON the screen. The design of agents, especially of the expert, has been determined by this principle: the rule conditions `read' the problem state on the screen, the rule conclusions change the problem state display. This design principle produces a nice opportunistic expert-learner collaboration and makes this interaction inspectable and tunable by the tutor. Experimentation with users helped us to find problems related to this approach, namely the difficulty of interacting about goals. But they also gave us some insight for implementing mechanisms of social grounding between a human user and a machine. Social grounding is a mechanism of joint and reciprocal modelling between two partners, based on the negotiation of shared meanings. A limitation in our current architecture is that the rule variables are instantiated by screen objects without any ambiguity. A new research avenue would test interactions where the computer's concepts were not directly connected to the displayed objects (the learner concepts are neither), and where these mappings could be negotiated with the learner through various pointing mechanisms.