A Scalable Solution to Architecting Standalone Large-scale Multi-agent Systems for Real-time Interaction and Rendering

In response to the trend of personalizing computing resources, this paper proposes a novel software architecture for multi-agent systems (MAS) to support real-time interaction and rendering in highly complicated interactive circumstance of thousands of agents with only one computing device. The architecture addresses challenges arising from complex business logic demands, where individual agents may encompass multifaceted attributes and methods, risking the emergence of god classes. To mitigate this, I adopt the principle of separation of concerns by modeling each agent as a composition of multiple Aspect Entities, each encapsulating distinct functional aspects. The proposed architecture incorporates a number of Scions and a Rootstock. Each Scion handle specific agent functionalities by managing these Aspect Entities, while Rootstock facilitates interaction among Aspect Entities of every agent through Agent Bus, enabling coordinated task execution. I validate this architecture through GUSU, a real-time strategy (RTS) game engine under development. Experimental results demonstrate that the proposed architecture effectively supports real-time interaction (about 70ms latency) and graphics and sound rendering (about 15.51 FPS with Win32 GDI and about 44.96 FPS with Direct2D) for large-scale agent populations (2500+), achieving low-latency performance in standalone settings. By decentralizing agent responsibilities and optimizing resource allocation, this solution ensures maintainability and scalability, aligning with the demands of modern personalized computing environments.