2 Criteria for a Good Model of Agent Learning

2.2 Incrementality of the Learning Process

One aspect of learning that seems clear is that agents frequently do not perform exhaustive searches for the best internal model of something, but adapt existing models. A good example of how well this works as a representation of actual agent model-building is Lenat's [12] model of how mathematicians form conjectures in number theory. This may be formalised by putting a cost on the search for new models. Even though the cost of each search step may be low, the total cost can still be significant. Thus discarding a model and starting again can be very costly due to the large search space of possible models. In such circumstances, it is important to be able to improve models by means of a focused search in the space of possible models. This seems to us to be an appropriate policy to follow in model-building of all kinds. We therefore assume that such a policy will be followed by agents in the development of their internal models.

Modelling Learning as Modelling - 23 FEB 98

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

Generated with CERN WebMaker