2 Criteria for a Good Model of Agent Learning

2.5 The importance of the form of the agent's models

Only some subspaces will be expressible in the language, the agent may be forced to approximate the actual subspace of possibilities with an expressible one.
Some subspaces will have several corresponding expressions, some of which will be far more efficient to use than others.
Although the agent will often know what the space of possibilities in some theoretical sense, she will often not know what expressions are necessary to describe a particular subspace, thus for all practical purposes the space of possibilities is unknown and unknowable.
A global search of possible subspaces is impractical, because the agent has to do this by searching through possible expressions in her language. This involves a different kind of search than that of paramaterising a known agent model.
If there is any significant cost associated with building the agent model expressions then some sort of incremental, path-dependent development of agent models is inevitable.
The agent may come to believe in inadequate or even partially inconsistent expressions. These would be less likely to be successful models, but are not automatically ruled out (depending on the agent language).

Generated with CERN WebMaker