StarCraft strategy learning refinement using replay snapshotting
Štefan Krištofík, Michaela Hanková · Annals of Computer Science and Information Systems · 2025
We propose a new replay snapshotting (RS) technique for strategy learning from past matches in real-time strategy game StarCraft: Brood War (SCBW).It allows for more precise understanding of particular strategy aspects by sampling the state of selected game features at important checkpoints during a match.We use RS to extract and refine a set of strategies from a large replay dataset STARDATA.To validate our approach in a competitive environment, we implement an AI agent for SCBW.It is able to perform the extracted strategy set against opponents in the BASIL Ladder competition.The agent consistently achieves rank C with 56 % win rate which is a significant improvement over our previous approaches.