Machine Learning in StarCraft II - Lowering the Difficulty Threshold of Starting From Scratch

Abstract

Artificial intelligence research is currently a hot topic within many industries. In terms of research, games such as StarCraft II provide a good testing ground due to its accessibility. However, getting started can still be more difficult than it should be. This paper aims to facilitate the development of a machine learning agent for StarCraft II by designing tools for data collection, making a simple API built on top of PySC2 to facilitate interaction with the game and by analyzing a few different types of artificial neural networks with respect to StarCraft II. It is concluded that defining reward functions for reinforcement learning can give rise to unexpected behaviors. A further conclusion is that convolutional neural networks tend to be more resource intensive than non-convolutional networks and that they are thus less suited for anyone without access to large computational power. Lastly, a network is trained on collected data to continuously predict the win chance for players in a StarCraft II match. Unfortunately the network does not become successful in its task, likely in part due to the simplicity of the network.