A Framework for Model-Based Adaptive Training

KBS Analysis and Design

In the AI literature there is a considerable amount of work focused on KBS analysis and design which can be extended to penetrate the ITS field. The MOBIT project has made tentative steps forward. To build on this work and bring together other KBS methods, Section 2.4.1 starts with a discussion of knowledge engineering. This is followed in Section 2.4.2 with an introduction into the foundation for MOBAT task decomposition methods.

Knowledge Engineering

In Europe, KADS is widely known research work in AI/KBS analysis and design methodology [Tansley & Hayball 1993]. ESPRIT funding for the first KADS project began in 1983. Progress in KADS-II is concentrating on the baseline methodology, called CommonKADS. The project has produced an innovative (but fragile) workbench which is intended to support construction of domain models with a library of re-usable components. In CommonKADS a series of knowledge engineering models for Organisation, Task, Agent, Expertise, Communication and Design are used to represent different aspects of the knowledge engineering process. These knowledge engineering models drive the analysis and design processes, and support maintenance by providing a persistent knowledge representation. Specified relationships between the KADS knowledge engineering models help the developer in identifying the impact of changes.

A set of knowledge engineering models with their specified inter-relationships can guide the design of both the expert agent and trainer agent. Guidelines on how to build KBS’s and ITS’s are still incomplete. Despite its limitations (e.g., incomplete support libraries) and perhaps some controversial elements, KADS has the advantage of being principled. It has reusable knowledge engineering models and a tool kit, and it is emerging as a reference standard (at least in Europe). KADS appears to be robust enough to be used in some industrial applications.

The word “model” is used here in relation to KADS knowledge engineering models.

Approaches in software engineering usually involve designing programs as a series of modules, which are easily understood and have a limited set of relationships with other modules. A componential framework that stresses modularity and an analysis of the pragmatic constraints on the task is discussed in [Steels 1990]. The componential framework attempts to synthesise four key elements: inference structures, deep expert systems, problem solving methods and generic tasks. Each of these elements focuses on one aspect of expertise and problem solving: the inference pattern, the domain models, the problem solving methods and the task features. The componential framework suggests cataloguing components for a knowledge engineering handbook that relates task features with expert systems.

A well-known KBS approach with loosely coupled knowledge-based modules is the blackboard model. A survey of blackboard KBS systems is documented in [Nii 1986]. A system called Capra [Fernandez-Castra et al. 1993] provides a good example of an ITS with a blackboard architecture. The blackboard model of problem solving is a highly structured special case of opportunistic problem solving. The blackboard model is a conceptual entity which is usually described as consisting of three major components: (1) the knowledge needed to solve the problem is partitioned into knowledge sourceswhich are kept separate and independent; (2) the problem solving state data are kept in a global database called the blackboard; and (3) the control component which can be kept in the blackboard, the knowledge sources (i.e., knowledge sources with their own control component can respond opportunistically to changes in the blackboard), in a separate control module or in some combination. This approach can be applied in an ITS system to separate instructional knowledge to control use of the knowledge sources and simplify incremental construction of new types of knowledge sources. In an ITS system a separate control module can represent the tutor.

A modular framework can have any type of knowledge source added a step at a time. Although the MOBAT specification framework does not assume a training system with a blackboard approach, the training architecture (see Figure 1-1) is modular with a generic trainer and one or more expert agents. The use of multiple (loosely coupled) expert agents allows a different knowledge base to be added a step at a time.


Flower Show
© | | Sitemap