[Next] [Previous] [Up] [Top] [Contents]
6.1 Example strategies for agent model development
Specialization of a model always increases the volume of the universal model and, so, is to be avoided where possible. Adding conditions of application always specializes an existing model. So, too, does increasing the scope of the model by adding variables. Anything which reduces the domain of a special model is obviously a specialization. It seems sensible to recognize any volume-increasing adaptation of a model as a specialization.
Two natural means of specializing models are by adding variables and by combining them in ways which facilitate domain-reduction.
When an agent adds a new variable to his model, this does not change the space of possibilities but it does change the size and perhaps the complexity of the model description. When the new variable is defined in terms of existing variables, it is a purely formal device in the model description, as whenever it occurs, the original reference to the variables it was defined by could be substituted instead. Placing conditions on this new variable is equivalent to a restrictive condition on the possible combinations of the environmental variables.
Take an example where there are two input variables v and w and one output variable p. If there is an existing model of the form p=g(x,y) for some relation g then the agent might try a model of the form p=f(v) where 
using some a priori information about the dimensionality of variables. The agent has thus decomposed the model into two separate parts to reduce the information considered. In effect, new conditions are added by specifying some relationship between the outputs of the models and combinations of the inputs to the model. This is a clear specialization of the original model (It is also an unambiguous simplification of the model as will be seen in section 7).
Model specialization takes the following form:
- Combine existing variables into a new variable in such a way that the number of dimensions of the new variable is less than the number of dimensions of the component variables. The only exception is combinations of pure numbers which are themselves pure numbers.
- Assess whether the changes in or values of the new variable discriminate between cases in which any model is disconfirmed and cases in which it is not.
- If it is found to have some discriminatory power, specialize the model by adding the appropriate statement to its conditions.
Modelling Learning as Modelling - 23 FEB 98
[Next] [Previous] [Up] [Top] [Contents]
Generated with CERN WebMaker