A Framework for Model-Based Adaptive Training

Foundation - Computer Models of Learning

Can learning for both the computer model and the trainee be a gradual move from a knowledge-lean to a knowledge-intensive situation? A good human trainer must have an understanding of processes of learning and development within the trainee. Similarly, an intelligent computer-based trainer requires well-understood processes of learning and development. An ITS should itself embody principles of learning.

This principle has extensively been applied in ITS’s produced by Anderson with a theory of cognition based upon a production system model of learning [Anderson et al. 1990]. Learning mechanisms in Anderson’s ACT* theory involve:

  1. declarative recording – information from the environment is deposited in working memory in a declarative manner;
  2. application of declarative knowledge to new situations (i.e., situations for which productions do not exist) by means of analogy and pure search;
  3. strengthening - in working memory and production memory the likelihood of knowledge being selected again is strengthened each time it is used; and,
  4. knowledge compilation – replacement of an extended computation into a single production rule. There are two forms for knowledge compilation: proceduralisation, where a general piece of knowledge is transformed into a specific production rule, and composition, where several rules are combined into one.

Anderson and his colleagues used this learning theory to build tutors for Lisp programming, for proof generation in high-school geometry, and for solving algebraic manipulation and word problems. These domains were selected because they involve the acquisition of well-defined skills [Anderson et al. 1990].

In the ACT* based systems two models for tutoring are considered:

  1. performance models – how students execute the skills that are to be tutored; and,
  2. learning models – how skills are required.

Performance models consist of correct and incorrect rules which are used with a model-tracing paradigm. Model tracing is a diagnostic method for the tutor system to detect deviations from the ideal student. The learning model consists of a set of assumptions about how the student’s knowledge state changes after each problem solving step. This is used in knowledge tracing to track the changes in the student’s knowledge across problems.

The model tracing paradigm is considered a very effective approach for limited subject domains. A problem with Anderson’s approach for tutoring systems, as pointed out by [Hill & Johnson 1993], is that model tracing has been designed for static tasks. Model tracing is best suited for subject domains where there are no unpredicted influences from the external environment and when there is only one way to do a task. In industrial (dynamic) environments a wide range of (problem solving) methods for accomplishing a task may be acceptable.

Another cognitive model is SOAR, an architecture that is based upon a theory of human cognition. In SOAR learning consists of the creation of additional rules, by hand coding or by a (single) learning mechanism that automatically chunks the results of successful goals [Rosenbloom 1983]. The chunking mechanism creates new production rules that allow the system to directly perform actions that originally required problem solving in subgoals.

The conditions of a chunked rule test those aspects of the task that were relevant to satisfying the goal, while its actions generate the information that actually satisfied the goal. New rules form part of search control when they deal with the selection among objects, or they form part of operator implementation when they are chunks for goals dealing with problematic operators. SOAR is driven by the goals automatically created to deal with impasses in its performance and chunking works for all goals. The chunking mechanism is applicable to all aspects of SOAR’s problem-solving behaviour [Rosenbloom & Assman 1990].

To have a task formulated in SOAR is to have a problem space and the ability to recognise when a state satisfies the goal of the task; that is, is a desired state. The default behaviour for SOAR – when it has no search-control knowledge at all – is to search in this problem space until it reaches a desired state. Search control knowledge can be added to any problem space in the form of new object preferences. or additional information that leads to new preferences.

As more knowledge is added, the problem solving becomes more and more constrained until finally search is totally eliminated. This is the basic device in SOAR to move towards a knowledge-intensive system. Each addition occurs simply by adding rules in the form of productions.

Theoretically, SOAR is able to move continuously from a knowledge-free solver (the default), through the weak methods to a knowledge-intensive system. It is possible to eliminate entire subspaces if their function can be realised by search-control knowledge in their superspace.

For instance, if a subspace is to gather information for selecting an operator, then that information might be encodable as search control in the higher space. Similarly, if a subspace is to apply an operator, then specific instances of that operator might be carried out directly by rules in the higher space.

Flower Show
© | | Sitemap