Web-Based Interface for Data Labeling in StarCraft

In-Chang Baek, Kyung-Joong Kim

Abstract: Recently, StarCraft AI has been very actively researched, largely via analysis of human replay data. However, such data are difficult to evaluate visually because they represent information from a limited environment, that of the game client. To solve this problem, we created an environment in which game screens are displayed on the web, allowing game progression to be evaluated at a glance. This allows the performance of more diverse and efficient experiments than conventional human testing. We show that human players label macro decisions (e.g., main force operations) during supervised StarCraft learning using a web-based interface.