Behavior Learning in Minority Games

Abstract

The Minority Game (MG) is a simple model for the collective behavior of agents in an idealized situation where they have to compete through adaptation for a finite resource. It has been regarded as an interesting complex dynamical disordered system from a statistical mechanics point of view. In this paper we have investigated the problem of learning the behaviors in the minority game. We assume the existence of one agent who can learn from other behaviors. We consider two scenarios in this research: (1) Given an environment with complete information, i.e., all records of agents' choices are known to public. The intelligent can use a Decision Tree to learn the patterns of other agents and make predictions. (2) If we only know the data of collective behaviors, we assume the data are generated from combining the behaviors of variant groups of agents. The intelligent can use a Genetic Algorithm to optimize these group parameters in order to get the best guess of the original system. The experimental results show that, in this configuration of MG in both environments with complete information and incomplete information, the intelligent can learn from history data and predict which side is the minority.