Real-time strategy (RTS) games demand agents to manage economy, production, construction, base defense, unit organization, and attack timing over extended matches. Prior work applied large language models to read textual game states and output high-level plans, but inference latency often causes missed tactical events, and the complexity of full matches forces heavy reliance on manually crafted experience prompts, limiting continuous learning from past games.
We introduce STRATA, a role‑aligned hierarchical system with cross‑game self‑learning for Red Alert. Decision making is split into three agents: the Strategic Agent (SA) issues global directives, the Logistics Agent (LA) handles resource allocation and building, and the Tactical Agent (TA) carries out unit micro‑management and attacks. After each match, a Review Agent (RA) extracts candidate experience from game traces, validates and revises it using evidence from subsequent matches, and compresses verified experience into concise experience cards for SA retrieval.
The experience‑card creation pipeline consists of: (1) extracting key state‑action pairs from full match logs; (2) testing whether these pairs improve win rates in later matches; (3) clustering and abstracting similar experiences into quickly searchable cards.
In a fixed scenario, the win rate without experience cards is about 30%, while incorporating learned cards raises it to 100%. Sequential learning against AI opponents with different play styles yields distinct long‑term strategic experiences—defensive, aggressive, and economy‑focused—demonstrating significant strategic diversity.
These findings show that a role‑aligned hierarchical agent architecture combined with post‑match self‑learning can maintain real‑time responsiveness while continuously evolving strategies, offering a scalable path toward general AI for RTS games.
Review