Learning to Identify Rush Strategies in StarCraft

摘要

This paper examines strategies used in StarCraft Ⅱ, a realtime strategy (RTS) game in which two opponents compete in a battlefield context. The RTS genre requires players to make effective strategic decisions. How players execute the selected strategies affects the game result. We propose a method to automatically classify strategies as rush or non-rush strategies using support vector machines (SVMs). We collected game replay data from an online StarCraft Ⅱ community and focused on high-level players to design the proposed classifier by evaluating four feature functions: (ⅰ) the upper bound of variance in time series for the numbers of workers, (ⅱ) the upper bound of the numbers of workers at a specific time, (ⅲ) the lower bound of the start time to build a second base, and (ⅳ) the upper bound of the start time to build a specific building. By evaluating these features, we obtained the parameters combinations required to design and construct the proposed SVM-based rush identifier. Then we implemented our findings into a StarCraft: Brood War (StarCraft Ⅰ) agent to demonstrate the effectiveness of the proposed method in a real-time game environment.