A Parallel FIPA Architecture Based on GPU for Games and Real Time Simulations

The research introduces a novel architecture for implementing Multi-Agent Systems (MAS) based on the Foundation for Intelligent Physical Agents (FIPA) standard using Graphics Processing Unit (GPU) computing. The primary thesis is that while traditional CPU-based agent systems struggle with the computational demands of massive crowd simulations in games and virtual environments, mapping these systems to a Single Instruction, Multiple Data (SIMD) GPU architecture significantly improves scalability and performance.

The methodology involves developing a GPU-based framework that maps agent behaviors to CUDA threads, utilizing a container-based approach to manage agent lifecycles. The researchers tested this architecture by implementing a pathfinding agent using the A algorithm. Performance was evaluated by comparing execution times on an Intel Core i7 CPU against a GeForce GTX 580 GPU across scenarios ranging from 100 to 1,000,000 agents. The findings demonstrate that while CPU execution times increase linearly and become prohibitive as the number of agents grows, the GPU architecture maintains consistent, real-time performance, effectively handling up to two orders of magnitude more agents.

The study concludes that GPU-based FIPA architectures are highly effective for massive, low-communication agent simulations. However, the researchers note significant limitations, specifically regarding the difficulty of implementing heterogeneous agent behaviors and inter-agent communication due to the constraints of the SIMD paradigm. Consequently, the proposed architecture is currently best suited for scenarios requiring high-density, uniform agent behaviors. Future work aims to develop an abstraction layer to simplify GPU programming for developers and to address communication and heterogeneity challenges within complex environments.