Previous Page Table of Contents Next Page

First: demonstrating the feasibility of simulating the skill

This approach focuses on a relatively static version of the skill or skill set. Passing a Turing test is a reasonable goal for this approach. The Turing test requires that observers cannot readily tell whether the behaviour originates from a real person or from a computer simulation. We look for algorithms that the computer or robot can use to produce not just a single skilled behaviour but a set of behaviours that are appropriate in different contexts and circumstances.

The robot will try to recognize the context and produce physical behaviour that is similar to that of a human. The computer will start with a symbolic description of the context and produce a verbal (or mathematical) specification for the behaviour. The computer, therefore, contains the computer and extends its capabilities to be more life-like by including perception and physical action. The robot is more labour-intensice and expensive to produce.

This approach has been quite successful but is rather limited in supporting skill engineering. We have programs that play chess very well (e.g. IBM's Deep Blue), but they do not simulate how humans play chess, and they cannot teach chess by diagnosing problems and helping to improve the skill.

The approach supports predictions that help control skilled behaviour by adjusting the context, i.e. the input. The approach is not sufficient for working with skill development since it simulates a static version if the skill, i.e. a skill that does not improve with experience, it does not learn.

Usability of the approach for skill engineering

There is a variety of such models, but only a limited amount of experience in using them for skill engineering. We start with the assumption that we have a working simulation, and go through the questions we are trying to address with the simulation.

  1. Does the model simulate the skill we are working with, say for teaching chess?
  2. Can the simulation be customized to different individuals to represent their level and version of the skill?
  3. Can an appropriate range of input situations be simulated, so that the model can deal with the context in which the target individual is applying the skill?
  4. Is the simulation output in sufficient detail so that it can mimic the skilled behaviour of the individual?
  5. Individuals have variations and flaws in their skilled behaviours. Can these be simulated in repeated run?
  6. Is the skill mechanism and information processing sufficiently transparent so that it can help diagnose and pinpoint why the individual illustrates a given variation in skilled behaviour?
  7. Can we use the model when customized to a given individual to indicate the range of behaviours that the individual might exhibit? In other words can we use the model the show whether the individual has the appropriate set of skills for the job to be done?