A Framework for Model-Based Adaptive Training

The Use of Multiple Models in a Training Framework

Methods are presented in this section which provide a consistent way in how the trainer agent is using the multiple modelling dimensions within the MOBAT specification framework. As introduced with Proto-MOBAT, both expert agents and a generic trainer agent are specified by using problem spaces, task primitives and methods for proposing and applying tasks. The idea of separating subject expertise and teaching expertise is not new.

For example, in Clancey’s GUIDON system, teaching knowledge is encoded by about 200 rules called tutoring rules or t-rules [Clancey 1986]. As reported by Clancey, these rules are encoded like a traditional expert system (i.e., a “first-generation” expert system) in a way that lacks theoretical foundation. Arbitrary strategies are encoded by GUIDON’s t-rules. Also, GUIDON assumes a particular problem solving method in the subject domain.

The problem solving method assumed is a diagnostic classification method motivated by the MYCIN medical expert system. Building on the GUIDON idea, the MOBAT specification for the trainer agent is based on not just one, but an explicit range of methods for switching models and adjusting models in the subject domain. The MOBAT design specification for the trainer agent is also resulting in tutoring rules, but these rules are organised by a set of trainer tasks (and trainer task primitives) and a library of trainer methods.

In Workmanship-MOBAT, all of the trainer methods include a “propose” step. This is the step that enables the SOAR-like model switching mechanism as proposed in Section 3.4. When a trainer task suggests the switching of models or adjustment of a model, then the propose step enables the evaluation of the most appropriate decision for a given situation.

Trainer methods can be used to explicitly evaluate different model switching suggestions. If no knowledge is available to distinguish among the suggested actions, then the fixed SOAR preference mechanism is used to select the next trainer decision from (any) competing alternatives.

The specification and realisation of the trainer agent is with instructional problem spaces, trainer tasks and trainer methods. The trainer agent uses didactic and diagnostic problem spaces which vary the seven domain model properties to deliver training units for effective training.

Within the Workmanship-MOBAT application, the didactic tactician is used to select and switch to the appropriate domain model using the (model) dimensions of scope, generality and perspicuity. Depending on trainee performance, the diagnostic tactician is used to adjust a model using the precision, accuracy and uncertainty model dimensions.

The following two sections present the didactic and diagnostic problem spaces in terms of trainer tasks and trainer methods.

© | | Sitemap