A Framework for Model-Based Adaptive Training

The Knowledge Level

The knowledge level is about defining the problem solving arena and how knowledge is used. In this work, the knowledge level consists of a set of problem spaces with tasks, methods and a model description. The trainer agent provides reasoning about a trainee model and the expert agent provides reasoning about a domain model.

For the expert agent, the knowledge level tasks are defined by adopting a generic task classification. Of course, tasks should not be directly linked to a particular way to do a task. There may be many ways to accomplish a task. The problem solving method is the activity to accomplish a task. It is not a single action, but a set of actions (i.e., problem solving steps) organised in a structure.

For each task at least one problem solving method needs to be identified. Methods can efficiently be applied to a fairly broad set of tasks, or each task can have one or more domain specific methods. Below is a summary for a generic method which has been used extensively in the MOBAT application experiments.

The MOBAT specification of problem solving methods is based on the premise that there are several ways of doing a task. These methods can be specified and stored in a problem solving method library. The primary methods involve a variety of knowledge role combinations (see Section 2.5.5).

A typical problem solving method together with its knowledge roles is described below. This method is derived from work done on RIME [van de Brug, Bachant & McDermott 1986]. As the MOBAT framework provides a detailed task decomposition specification, and to provide a closer link between knowledge level and symbol level design, the use of ‘operators’ in RIME to control problem solving is replaced by ‘tasks’.

  • Step 1: Propose candidate tasks (propose knowledge role). An expert agent's first step in a new problem-space is to propose tasks that are relevant to the current situation, and to discard those whose pre-conditions are not satisfied. At the end of this step, all of the tasks (and sub-tasks) that could plausibly be done are available for consideration.
  • Step 2: Eliminate inappropriate and obviously inferior candidates (reject knowledge role). The expert agent’s second step is to try to prune some of the candidate tasks. If there are candidate tasks whose preconditions are all satisfied, any tasks whose preconditions are not all satisfied are pruned. If there are candidate tasks whose preference class is lower than the preference class of other candidate tasks, those tasks in the lower preference class are pruned.
  • Step 3: Evaluate the remaining candidates (evaluate and select knowledge roles). During the third step, the expert agent compares the candidate tasks that remain after the second step with one another. Circumstances that suggest that one of the tasks is less appropriate than another, results in the elimination of the less appropriate task. At the end of this step, all of the evidence that the expert agent has that allows it to discriminate among the candidate tasks has been taken into account. If more than one candidate remains, the expert agent selects one of the candidates at random
  • Step 4: Perform the actions associated with a selected task (apply knowledge role). The expert agent's fourth step is to apply the task selected in the previous step. If the task can be realised within the current problem-space, whatever actions are performed to realise this task are performed during this step. If the task is complex and can only be realised by invoking another problem-space, the problem-space in which the selected task can be realised is invoked.
  • Step 5: Iterate or Quit if there is nothing more to do (recognise-success and recognise-failure knowledge roles). In the fifth step, the expert agent looks for evidence that it has done all that can be done for now in the current problem-space. If it recognises that it has done everything it can, control returns to Step 5 in the parent problem-space. If the expert agent finds no success or failure knowledge, it is appropriate to iterate through the steps again, and so it goes to Step 1.

These knowledge roles can be combined to create different problem solving methods. Propose & apply is a straightforward linking of a situation with actions to be applied (i.e., a general inference engine determines the actions to apply). Propose, evaluate & apply will ensure an explicit (knowledge-based) evaluation of alternative actions.

For the MOBAT framework, a domain specific set of knowledge roles and a library of problem solving methods can be used (which is depending on the particular training domain). The expert agent is a domain specific component in the training system architecture. Either very subject specific or quite generic methods may be appropriate.

The important consideration in the MOBAT framework is to define explicit methods that can be recognised by the trainer agent to help the trainee to reflect on what happens during problem solving and select among alternative problem solving steps (both correct and incorrect steps).

Flower Show
© | | Sitemap