A Framework for Model-Based Adaptive Training

Task Decomposition

There is wide-spread confusion in the Knowledge Engineering literature about what is meant by a task, i.e. an aspect of an (application) problem description or a problem solving method. In MOBIT, task decomposition means classification of training objectives with domain specific tasks as opposed to the idea of generic problem solving (reasoning type) tasks and task specific architecture’s. In [Chandrasekaran 1986], six generic problem solving tasks are described as building block for the construction of KBS applications:

1. Hierarchical classification;
2. Hypothesis matching;
3. Knowledge-directed information passing;
4. Abductive assembly;
5. State abstraction; and,
6. Hierarchical design by plan selection and refinement.

The first 4 of these tasks are designed for diagnostic reasoning and the last 2 for planning systems.

Separating knowledge from its intended use can lead to a number of difficulties in KBS design and maintenance (e.g., sometimes control elements are implicitly embedded in the knowledge base). A motivation for designing generic problem solving tasks and task specific architectures is the view that the representation of knowledge should closely follow its intended use, since the form of knowledge is tied to its use, and that there are different architectures for different types of problem solving [Chandrasekaran 1986]. A motivation for domain specific tasks (application tasks) is to make it easier to specify the different ways to do a task and to make it easier for the domain expert(s) to directly participate in the design of the knowledge base.

The KADS methodology makes use of a generic task library. However, the difficulty in using this library is that it is incomplete. A generic task set that is considered complete from a systems engineering perspective is the QUIC task classification [Leitch & Gallanti 1992]. The QUIC project identifies five primitive tasks:

1. interpretation – transformation of observations into the adopted state representation;
2. prediction – generation of future states from the known or assumed current state;
3. identification – determination of unknown past states from known or assumed current states;
4. decision – derivation of conclusions from the assumed or augmented state; and.
5. execution – transformation of conclusions into actions that are carried out.

Figure 2- 1 QUIC Generic Task Classification [Leitch & Gallanti 1992]

The MOBAT specification framework has adopted the QUIC task decomposition for decomposing training objectives. This defines the tasks a trainee is expected to learn. Task decomposition is described further with application experiments in the following chapters.