A Framework for Model-Based Adaptive Training

Foundation - Model-based Approach

The application of Artificial Intelligence techniques in training systems is not new. The most common application of AI techniques is with expert systems representing empirical knowledge of associations or heuristics. The shallow reasoning of the ‘first generation’ approach to constructing expert systems is widely recognised and various approaches to removing the limitation of these early expert systems exist. Explicit models can be used to provide a more general reasoning system, independent of a particular (training) application. Unless indicated differently, the term model is used in a very general sense to mean any representation of the training problem that is constructed for the purpose of solving a problem, or set of problems.

Models are used as the basis for the problem solving process. These models are defined by the structure of the problem domain, how the domain operates and problem solving knowledge. Since multiple problem solving processes may be relevant for a particular application system, the use of multiple models may be needed. Several researchers in intelligent training systems have suggested that multiple models can provide a rich environment for intelligent training systems (e.g., [White & Frederiksen 1990, Sime 1994]). The MOBAT specification framework is based on multiple explicit models of cognitive processes and qualitative models of the physical device, product or process on which the training is to be performed. A cognitive modelling approach and a concept for identifying the scope of an appropriate area for model-based training is introduced with problem-spaces in Section 2.3.1.

Cognitive Modelling in Problem Spaces

Examples of cognitive modelling are the ACT* theory [Anderson et al. 1990] and the SOAR approach [Laird et al. 1987]. These approaches are candidate theories of human cognition therefore designing an ITS with it can show a close correspondence to human learning behaviour. In both theories learning involves the creation and enhancement of production rules in a problem solving process. Although Anderson’s approach has produced ITS’s which are considered to be successful, these general approaches tend to be rejected, not because they are wrong, but because their complexity makes it difficult for practical applications [VanLehn 1993]. To examine this further, aspects of SOAR research are reviewed for further analysis.

SOAR is a problem solving system based on formulating all problem-solving activity as attempts to satisfy goals via heuristic search in problem spaces [Laird et al. 1987]. A problem space consists of a set of states and a set of operators that transform one state into another. Starting from an initial state the problem solver applies a sequence of operators in an attempt to reach a state that satisfies the goal (called a desired state). Each goal has associated with it a problem space within which goal satisfaction is being attempted, a current state in that problem space, and an operator which is to be applied to the current state to yield a new state. The search proceeds via decisions that change the current problem space, state, or operator. A subgoal is attempted by selecting a problem space for it, with goal attainment interpreted as finding a desired state in that problem space. Should a decision be problematic in this new problem space, a new subgoal would be created to deal with it. The overall structure takes the form of a goal-subgoal hierarchy. Each new subgoal will have an associated problem space. SOAR generates a hierarchy of problem spaces, as well as a hierarchy of goals. The diversity of task domains is reflected in a diversity of problem spaces. Major tasks have a corresponding problem space, but so also do each of the various subtasks. For example, sub-problem-spaces for configuring a computer may be the placing of a module into a backplane or placing a backplane into a box [van de Brug, Bachant & McDermott 1985].

Since a purpose of learning is in solving problems, the internal model to be created from the environment is equivalent to a problem space [Newell & Simon 1972]. For all problem solving situations, a problem-space is the arena within which problem solving takes place. In a mature problem space (e.g., games & puzzles) extracting and mapping knowledge from the environment is only done once. The MOBAT specification framework for training applications involves mapping of a considerable amount of training knowledge into problem spaces. This is ongoing because the learning cannot be separated from the actual environment (which in industry is always changing). The hierarchy of problem spaces can be applied for dynamic training where the training objectives are the root nodes of a goal hierarchy. The SOAR problem solving approach can be compared to training situations where a training goal is formulated by comparing the current trainee’s expertise (“what is”) to the desirable state (“what should be”) or goal expertise.

In MOBIT, the scope of a domain model defines the physical extent of the system being modelled. By using problem-space representations for the scope of a problem, this modelling dimension is expanded in MOBAT to mean both the physical and conceptual extent of the system being modelled.
Flower Show
© | | Sitemap