A Framework for Model-Based Adaptive Training

Foundation - Problem Solving Methods

The CommonKADS modelling framework merges ideas from various approaches to knowledge modelling (e.g., Generic Tasks [Chandrasekaran 1986], role-limiting methods [McDermott 1988], Components of Expertise [Steels 1990]). Problem solving methods in KADS can be seen as part of the re-usable components combining domain, task and inference knowledge.

This section describes some problems with a computer configuration expert system called R1 (or sometimes XCON) and the underlying techniques of an improved methodology (called RIME) involving problem solving methods [van de Brug, Bachant & McDermott 1986]. This approach serves as the basis for the MOBAT problem solving methods for the expert agent and the control methods for the trainer agent.

In performing the configuration task, R1 takes as input a list of components a customer has ordered and produces as output a set of diagrams of the interrelationships among those components. The initial list of components may be incomplete (i.e., it may not be possible to configure a functional system with that set of components) and, if so, R1 must add appropriate components.

The number of possible component combinations is very large. The only reasonable approach to configuring the components is to construct (as opposed to select) an appropriate system. Generally, constructive tasks perform heuristic search, that is, a combinatorial search in which candidate partial solutions are constructed and their potentials evaluated. R1 can frequently avoid combinatorial search (and therefore avoids backtracking) by using small local searches for additional information at steps where there is ambiguity about what next action is most appropriate. In other words, local cues are ordinarily sufficient to drive R1 along a path to a solution.

The initial versions of R1 were based on an implicit Problem-Solving Method. R1's problem solving involves the selection of the next piece of knowledge to apply from among those associated with the currently active subtask. Generally, only a few possible actions are relevant at any given time. A piece of knowledge is considered relevant whenever the pattern defining its relevance can be instantiated by elements describing the current configuration state.

When more than one piece of knowledge is relevant, the problem solving method relies on very general heuristics, such as the recency of the elements and the specificity of each pattern, to determine which piece to apply. R1's problem solving can be characterised as follows:

Given that it is involved in some task, it will take whatever next action (i.e., apply whatever knowledge) is relevant; if more than one piece of knowledge is potentially relevant, the choice will be made on the basis of very general considerations; if there is no more knowledge relevant to the current task, R1's attention returns to the parent task; whenever R1 does not have enough information to confidently prefer one possible action to all other candidate actions, it does some local problem solving (e. g ., by invoking some information gathering subtask) until sufficient information has been collected [van de Brug, Bachant & McDermott 1986].

The Arduous Matter of Adding Knowledge

R1's implicit problem solving method does not provide expert configurers with clear guidelines about what knowledge they are expected to share. In particular, a person adding knowledge to R1 could use help in

  1. how to go about bounding the potential relevance of a piece of knowledge; and,
  2. how to determine which piece to apply when more than one is relevant.

The relevance of each piece of R1’s knowledge is defined by a pattern (i.e., a set of conditions): the pattern specifies, for some subtask, the circumstances under which the piece of knowledge can be applied. As R1 has no defined knowledge roles, the way relevant pieces of knowledge are chosen cannot be explicitly expressed. The problem solving method provides no vocabulary for an expert to describe the various roles knowledge will play in the performance of a task.

Knowledge has been represented in R1 in various ways; regularities, to the extent they exist, have gone unnoticed. One has to know R1 well to modify its behaviour in some desired fashion. Since R1 has so much knowledge, gaining such familiarity is time consuming. It is therefore hard to communicate to the variety of people adding knowledge to R1 what is required of them. R1’s problem solving method is just a problem solving inclination that has to be further specified by the knowledge uses.

Flower Show
© | | Sitemap