Unsupervised Hierarchical Clustering of Build Orders in a Real-Time Strategy Game

Currently, no artificial intelligence (AI) agent can beat a professional real-time strategy game player. Lack of effective opponent modeling limits an AI agent’s ability to adapt to new opponents or strategies. Opponent models provide an understanding of the opponent’s strategy and potential future actions. To date, opponent models have relied on handcrafted features and expert-defined strategies, which restricts AI agent opponent models to previously known and easily understood strategies. In this paper, we propose size-first hierarchic clustering to cluster players that employ similar strategies in a real-time strategy (RTS) game. We employ an unsupervised hierarchal clustering algorithm to cluster game build orders into strategy groups. To eliminate small outlying clusters, the hierarchal clustering algorithm was modified to first group the smallest cluster with its closest neighbor, i.e., size-first hierarchal clustering. In our analysis, we employ a previously developed dataset based on StarCraft: Brood War game replays. In our proposed approach, principal component analysis (PCA) is used to visualize player clusters, and the obtained PCA graphs show that the clusters are qualitatively distinct. We also demonstrate that a game’s outcome is marginally affected by both players’ clusters. In addition, we show that the opponent’s faction can be determined based on a player’s transition between clusters overtime. The novelty of our analysis is the lack of expert-defined features and an automated stopping condition to determine the appropriate number of clusters. Thus, the proposed approach is bias-free and applicable to any StarCraft-like RTS game.