Evolving an Artificial Real-Time Strategy Game Player

Abstract

Computer-controlled opponents in commercial computer games are often unintelligent, but even more commonly extremely predictable in their behaviour. This detracts from the challenge of the game when playing against computer-controlled opponents, and is the primary reason players consider playing against a computer-controlled opponent significantly less fun and challenging than playing against another human player.

Many machine learning systems have been shown to be effective in creating more reactive, intelligent-seeming computer game players; the question of whether these systems are useful in a commercial context has already been answered. In this Master's thesis, I investigate the process of utilising one of these systems in the context of developing commercial computer game AI. That is; I don't aim to answer the question of *whether* these systems should be used in a commercial context, but *how* they should be used.

Because I cannot investigate all machine learning systems in the context of all different computer game genres, this project is narrowed down to one specific instance: The primary goal of this project is to empirically discover which parameters and factors are important when employing an evolutionary process to create an Artificially Intelligent player for a Real-Time Strategy game.