Predicting the Opponent in the Google AI Challenge: Ants

Abstract

Learning to predict the behaviour of an opponent is useful in many tasks, ranging from computer games to multi-robot situations. This thesis investigates algorithms for doing so in The Google AI Challenge: Ants, Fall 2011 edition, or simply called Ants Challenge. The Ants Challenge features a world populated by artificial ants. Behavioural data of these bots is used by decision tree algorithms to predict the behaviour of these ants. Literature shows that decision trees are capable of doing such task. This thesis will test various representations and algorithms for predicting the opponent. Results of the experiments show that simple, reactive bots can be predicted with a percentage of correctly classified instances of up to 98.1%.