A Framework for Model-Based Adaptive Training

Model Switching with Preferences

For a set of model dimensions and the SOAR cognitive modelling approach, can a model switching method involving preferences be used for a principled approach to using multiple domain models (i.e., deciding when to switch models and which model to switch to)?

This section introduces how the SOAR methodology makes decisions for model selection and switching to different problem spaces. This is kept brief as the SOAR cognitive modelling approach, plus a methodology for using SOAR with explicit representations (i.e., the RIME methodology [van de Brug et. al. 1986]), is presented in the previous chapter.

In SOAR, the knowledge used to make decisions is called search control. Knowledge to control search is expressed in preferences. As long as the search-control knowledge is adequate for the decisions to be made, problem solving proceeds smoothly. However, SOAR often works in domains where its search-control knowledge is either inconsistent or incomplete.

When this happens, an impasse in problem solving occurs. SOAR responds by automatically creating a subgoal whose purpose is to obtain the knowledge which will resolve the impasse and allow the decision to be made. For example, if more than one operator can be applied to a state, and the available knowledge does not prefer one operator over the others, a subgoal will be created to find information leading to the selection of the appropriate one. Another example is when an operator is selected and its implementation requires problem solving. A subgoal is created to build the state that is the result of the operator.

SOAR uses a production-system architecture that admits parallel execution of all satisfied productions to realise search-control knowledge and to implement its simple operators. More complex operators are encoded as separate problem spaces that are chosen for the subgoals that arise when the operator they implement has been selected to apply.

Each production rule elaborates the current objects under consideration for a decision (e .g ., candidate operators or states) with knowledge about the objects, including preferences relative to other candidate objects. Each decision corresponds to an elementary step in the problem solving (so a count of the number of decisions is a good measure of the amount of problem solving performed). There is a fixed process which interprets these preferences and makes a decision.

One type of data, a preference, can be used to control behaviour of a problem-space [Laird et al. 1987] and to select and switch to other problem-spaces. Within the MOBAT framework, this mechanism is used in selecting and adjusting appropriate models for adapting to trainee needs.

Flower Show
© | | Sitemap