A Framework for Model-Based Adaptive Training

Generic Trainer Realisation

The expert agent(s) provide reasoning about domain models (or cased models) and the trainer agent provides reasoning about a trainee model. The training unit specification provides the initial inputs to a (dynamic) trainee model. The trainee model in Proto-MOBAT includes:

  1. trainee profile information with trainee learning mode, preferences and capabilities;
  2. a dynamic training goal hierarchy; and,
  3. diagnostic information.

The specification of a static trainee profile is presented in Section 7.6. The trainer agent is a generic expert in instruction techniques. The specification of such a trainer is similar to the domain expert where the capabilities are designed at the knowledge level and implementation is at the symbol level. Similar to the domain expert specification, the knowledge level for the trainer is organised in problem spaces. The trainer design in Proto-MOBAT is limited to just three problem spaces, comprising a performance monitor, didactic tactician and diagnostic tactician. The performance monitor contains the communication and control mechanisms between active training units and the training agents. The didactic tactician is the software component which selects and directs the execution of training plans so as to achieve training goals. The diagnostic tactician is the software component which takes appropriate action when there are trainee errors and classifies errors as slips, bugs and misconceptions.

Trainer Specification Levels

Figure 4-7 Trainer Specification Levels

For the Proto-MOBAT trainer agent, the task primitives are: to present,to dictate, to coach, to facilitate, to advise, to intervene, to remediate and to assess. The first four of these task primitives are assigned to the didactic tactician and the other four are used by the diagnostic tactician. The main tactical instructional methods are a problem-driven-method, example-driven-method and test-driven-method. For a discussion of trainer task primitives and methods see Section 6.5. As the trainer in the training system architecture is a generic component for a range of training domains, the appropriate trainer functionality can be selected from a library of trainer tasks and trainer methods. A similar approach to designing a generic tutoring engine has been reported for a system called GTE [Van Marcke 1990, Van Marcke & Vedelaar 1995]. GTE has a large and impressive instructional knowledge base using an extensive library with over 153 instructional tasks and more than 314 instructional methods. MOBAT applications are expected to be able to operate with fewer instructional methods by making use of generic (but still explicit) methods such as propose-evaluate-apply and its matching knowledge roles as described in Section 2.5. GTE does not place its instructional knowledge base in a framework to include simulation and different expert models. The instructional tasks and methods operating on different model types are a key aspect of the MOBAT framework. A further discussion of trainer tasks, trainer methods and recognition of seven generic modelling dimensions within the MOBAT framework is presented in chapter 6.

At the symbol level there are representations for skeleton training plans, training heuristics and training unit scripts. Skeleton training plans are specified from the trainer interface – see Figure 1-1 Training System Architecture (adapted from [Slater et al. 1994] ). Examples of skeleton training plans are presented in Figure 5-3 and Figure 6-2. Several researchers in ITS have suggested the implementation of tutoring heuristics in the form of rules or productions [Anderson et al. 1990, Clancey 1986, Hill & Johnson 1993]. These heuristics can be created with either general purpose programming languages or specialised modelling tools. Training unit scripts are a slot in training units which are executed by the trainer component. A script is the general control structure for a training unit. As training units within the training system architecture are autonomously executing units, each training unit is provided with its own control structure. At run time the trainer agent can monitor and intervene in a training session as needed. An alternative design to a script-based control structure is to provide each training unit with their own associated training unit manager module. Specialised scripting tools are suggested in the MOBIT project [MOBIT 1994, W7].
© | | Sitemap