2 Criteria for a Good Model of Agent Learning

2.4 The existence of different trade-offs in model search

The most useful model for an agent would be one which, with complete confidence, gave point forecasts of every target variable for every feasible set of values of the decision variables and yet was reasonably simple. The purpose of model adaptation is to move from the default position towards the best position. This involves the best possible increase in precision with the least possible complexity. A noble sentiment completely without operational content and one which is, of course, not always possible. The guides to model improvement, should be able to include:

The error (e.g. RMSE) of the model's predictions. Clearly, a model with less error is better.
The complexity of the model. A simpler model is better than a more complex one for an The cost of searching for a model.
The volume, a summary measure of precision and generality, described below.

Sometimes these guides will indicate opposing directions of agent model development. In these cases the trade-off decided upon will be different for different agents (and also sometimes the same agents in different situations). For example, for a firm in a highly competitive but slowly changing market an accurate and error-free model of pricing might be more important than the complexity of the model or the cost of its development, while a firm with distinctly limited resources (a start-up?) looking for an opportunity in a fast moving environment might settle for a less accurate and vaguer model.

Modelling Learning as Modelling - 23 FEB 98

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

Generated with CERN WebMaker