Real-time strategy games are notoriously difficult for AI because they require juggling many competing objectives over long matches. A new paper on arXiv introduces STRATA, a framework that breaks this problem down by assigning different roles to a hierarchy of agents. Each tier handles a distinct aspect of play, such as base defense or unit organization, allowing the system to coordinate economic growth, production, and attack timing as a whole.
The key innovation is the use of role-aligned tiered agents that learn through self-play rather than from labeled data or hand-crafted rules. This lets the system develop its own solutions to the coordination problem, adapting to the game's evolving demands. The authors position STRATA as a step beyond existing large-language-model-based approaches, which have struggled with the long-horizon, multi-objective nature of RTS titles.
While the abstract outlines the framework's design and motivation, it does not yet provide experimental results or comparisons against other methods. The paper's contribution at this stage is the architecture itself, which offers a new way to decompose complex strategic decision-making into learnable, role-specific subtasks.