MultiAgent Techniques for Simulation of Complex Scienti cSystemsT

In this paper we present a paradigm for simulating complex systems. Such systems, typically found in Scientiic and Engineering domains (simulating an engine, designing an airplane wing), involve multiple physical phenomena with complicated geometry. The computational techniques we propose use cooperating agents. Our focus in this paper will be on phenomenon modeled by PDEs, although much of the proposed methodology is equally applicable to any scientiic computing task. We subdivide the physical object into components of simple geometric shapes modeled by a single, homogeneous mathematical model for which a problem solving environment (PSE) exists. PSEs are viewed as agents which solve the PDE on each component independently. The interfaces between the components must have physical interface conditions satissed; mediator agents use relaxation techniques for this. We brieey describe an agent-based architecture and its implementation (SciAgents) for building simulation systems that implement this paradigm. This approach is naturally parallel and highly scalable; it is suitable for a wide variety of parallel and distributed high performance computing (HPC) architectures; it allows for the reuse and evolution of existing HPC software, and for a convenient abstraction of the solution process.