Improving StarCraft II Player League Prediction with Macro-Level Features

Yinheng Chen, Matthew Aitchison, Penny SweetserAustralasian Conference on Artificial Intelligence 2020 · OpenReview

Abstract

Accurate player skill modelling is an important but challenging task in Real-Time Strategy Games. Previous efforts have relied strongly on micromanagement features, such as Actions Per Minute, producing limited results. In this paper, we present an improved player skill classifier for StarCraft II that predicts, from a replay, a player’s exact league at 61.7% accuracy, or within one league at 94.5%, outperforming the previous state of the art of 47.3%. Unlike previous classifiers, our classifier makes use of a macro-level measure of economic performance, called Spending Quotient, which we demonstrate to be an important part of accurately predicting player skill levels.
BibTeX
@inproceedings{chen2020improving,
  author = {Yinheng Chen and Matthew Aitchison and Penny Sweetser},
  title = {{Improving StarCraft II Player League Prediction with Macro-Level Features}},
  booktitle = {Australasian Conference on Artificial Intelligence},
  year = {2020},
  url = {https://openreview.net/forum?id=Ygnan0GSjpg},
}

Loading related content...

Figure