Inferring Micro-Rules from Macro-Behavior in the Minority Game

Wednesday, November 5, 2003 - 9:20am - 9:55am
Keller 3-180
Alexis Arias (Icosystem Corporation)
In many real world applications of ABMs, enhancing the predictive power of the model is paramount. This requires extensive knowledge of the behavioral rules that govern the agents' actions. However, it is common to face situations where direct information, or agreement among domain experts, regarding these rules is lacking, and only output sample data (usually at an aggregate level) is available. Under these circumstances, it is important to understand whether the behavioral rules can be identified from the observable data. The identification of micro-rules from macro- behavior in ABMs when there are nonlinear interactions between agents is the subject of this presentation.

I will present results from an ongoing project designed to study the possibility of estimating individual behavioral rules in the Minority Game using sample data that exhibits different levels of aggregation. The analysis concentrates on the small sample properties of the Maximum Likelihood Estimator of individual rules. We consider two models that pose different challenges with regard to estimation and three data scenarios: panel data of individuals' actions, time series of the number of individuals in the minority, and time series of the action taken by the minority. For each scenario we study the evolution of the estimation error as the number of individuals and the size of the time series increases. In addition, we analyze the effect on the estimation error of introducing certain restrictions on the model and on the information available a priori to the modeler.