Modelling Learning as Modelling
This paper is intended as a first statement that our approach to modelling learning and expectations formation is formally sound, practically relevant and, within the field of simulation modelling, describes behaviour which can achieve performance no worse than standard optimization algorithms or sound game theoretic strategies. There are a number of directions in which to pursue the implied research programme.
Because of their simplicity and the small number of observed data points, we have not assumed that agents estimate the parameters of their models statistically. A natural extension of this research is to apply it to more elaborate models which would allow for this assumption.
The strength of the approach described here is that it provides a formal description of learning as modelling which, in simple examples, performs better than textbook strategies. Learning as modelling offers a trade-off between being computationally inexpensive or parsimonious in its data requirements. The less data the agent uses, the more inventive must be his modelling. We would expect increased inventiveness to be associated with more elaborate and computationally expensive metarulebases. As any economist will recognize immediately, the rational agent will always reduce both computational resources and data requirements if these can be accomplished simultaneously without loss of relevant forecasting accuracy. Consequently, efficient decision-making will involve this trade-off. Simulations under a variety of alternative assumptions will inform both theoretical development and the analysis of practical alternatives in the organization of decision-making in conditions of complexity and uncertainty where the trade-off between computation and data acquisition is binding.
Generated with CERN WebMaker