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

7.1 The General Framework

7.1.3 The semantics of L


There will be a mapping , which will reflect the natural semantics of L in the space of possibilities, PS. Each member, l, of L refers to a subspace, m(l) (this, after all, is the point of such models). This mapping is usually not a 1-1 mapping, in that not every possible subspace of PS is mapped to by a member of L.

As discussed in section 2.5 this allows for the possibility of genuine surprise, when the agent faces some behaviour of its environment which it could not model. It is maybe this syntactic/semantic split that most differentiates this approach from a more traditional economic approach, where is sometimes assumed that behaviour of the economy and the model the agent uses are basically the same.

In general allowed models of L, will be only those that are consistent with the given a piori knowledge that the agent has, i.e. those which refer to subspaces of the space referred to by this knowledge. It is usual for the syntax of L, to be designed so that all expressions in L correspond to allowable models.

The indicators for the agent (utility, pain, profit etc.), are functions of the actual resulting position in PS, from which the agents goals may either be specified (or even inferred by the agent). This goal also corresponds to a subspace of PS.

Finally the agents knowledge of past behaviour can be represented by a series of data points in PS.

The general structure is illustrated in figure 4


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

Generated with CERN WebMaker