[Next] [Previous] [Up] [Top] [Contents]

7 Towards a Specification of a Framework for Modelling Economic Learning by Modelling

7.4 Model search

In general the first task for the agent at time t is to find a model , such that:

  1. (the model agrees with the a priori information);

  2. is as small as possible (the model agrees with the past data);

  3. is as small as possible, (the model is as predictive as possible);

  4. is as small as possible, where here is a suitably defined measure of syntactic complexity (e.g. the depth of the expression), this is closely linked with the cost of using the final model to choose an action;

  5. The cost of finding is as small as possible (the agent generates and tests as few candidate models as possible). The cost at this stage comes from the number of candidate models evaluated against past data;

These conditions on a suitable model are sometimes in conflict. For example an agent may be faced with a choice of a much simpler but marginally more distant agent model and a much more complex but slightly more accurate one. In general the order of precedence of these conditions will vary for different agents. For some agents the precision of the model and the agreement with the data will be more important than the complexity of the model description, for others a simple model with small computational cost would be preferable.

Modelling Learning as Modelling - 23 FEB 98
[Next] [Previous] [Up] [Top] [Contents]

Generated with CERN WebMaker