Emmanuel Hermelin, candidate for the position of MCF Polytech (ex-ESSTIN), will speak during the D3 seminar Thursday, February 15 at 10:30 in room C103. He did a PhD at LIRMM and is currently a research engineer at LIG.
Title: Modeling and implementation of multi-agent simulations on massively parallel architectures.
Abstract: Multi-agent simulation is a relevant solution for the engineering and study of complex systems in many fields (artificial life, biology, economics, etc.). However, it sometimes requires a lot of computing resources, which represents a major technological lock that limits the study possibilities of the models envisaged (scalability, expressivity of proposed models, real-time interaction, etc.).
Among the technologies available for High Performance Computing (HPC), General Purpose Computing (GPGPU) is the use of massively parallel graphics architectures (GPUs) as a computational accelerator. However, while many areas benefit from the GPGPU’s performance (meteorology, aerodynamic calculations, molecular modeling, finance, etc.), it is little used in the context of multi-agent simulation. In fact, the GPGPU is accompanied by a very specific development context that requires a deep and non-trivial transformation of multi-agent models. Thus, despite the existence of pioneering work that demonstrates the interest of the GPGPU, this difficulty explains the lack of enthusiasm of the multi-agent community for the GPGPU.
We show that, among the works that aim to facilitate the use of GPGPU in an agent context, most do it through a transparent use of this technology. However, this approach requires the abstraction of a number of parts of the model, which greatly limits the scope of the proposed solutions. To overcome this problem, and contrary to existing solutions, we propose to use a hybrid approach (the execution of the simulation is shared between the processor and the graphics card) which emphasizes accessibility and reusability through a modeling that allows a direct and facilitated use of GPU programming. More specifically, this approach is based on a design principle, called GPU delegation of agent perceptions, which consists of reifying part of the calculations performed in agent behavior in new structures (e.g., in the environment). This is to spread the complexity of the code and to modularize its implementation. The study of this principle as well as the different experiments carried out shows the interest of this approach as well from the conceptual point of view as from the point of view of the performances. This is why we have generalized this approach in the form of a method of modeling and implementing multi-agent simulations specifically adapted to the use of massively parallel architectures.
Subsequently, we have shown that the use of GPGPU in the domain of multi-agent simulations via this method, opened up many possibilities in terms of the resulting complexity of agent behaviors and / or the dynamics of these simulated systems. And this by using very simple multi-agent models and especially without adding artifices and / or simulation bias.